research paper about weather forecast

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 05 July 2023

Accurate medium-range global weather forecasting with 3D neural networks

• Kaifeng Bi   ORCID: orcid.org/0009-0008-7872-1197 1 ,
• Lingxi Xie   ORCID: orcid.org/0000-0003-4831-9451 1 ,
• Hengheng Zhang   ORCID: orcid.org/0009-0006-3241-6862 1 ,
• Xin Chen   ORCID: orcid.org/0000-0003-3686-6212 1 ,
• Xiaotao Gu   ORCID: orcid.org/0009-0002-7896-4750 1 &
• Qi Tian   ORCID: orcid.org/0000-0002-7252-5047 1  

Nature volume  619 ,  pages 533–538 ( 2023 ) Cite this article

215k Accesses

227 Citations

1723 Altmetric

Metrics details

• Atmospheric dynamics
• Computer science

An Author Correction to this article was published on 14 September 2023

This article has been updated

Weather forecasting is important for science and society. At present, the most accurate forecast system is the numerical weather prediction (NWP) method, which represents atmospheric states as discretized grids and numerically solves partial differential equations that describe the transition between those states 1 . However, this procedure is computationally expensive. Recently, artificial-intelligence-based methods 2 have shown potential in accelerating weather forecasting by orders of magnitude, but the forecast accuracy is still significantly lower than that of NWP methods. Here we introduce an artificial-intelligence-based method for accurate, medium-range global weather forecasting. We show that three-dimensional deep networks equipped with Earth-specific priors are effective at dealing with complex patterns in weather data, and that a hierarchical temporal aggregation strategy reduces accumulation errors in medium-range forecasting. Trained on 39 years of global data, our program, Pangu-Weather, obtains stronger deterministic forecast results on reanalysis data in all tested variables when compared with the world’s best NWP system, the operational integrated forecasting system of the European Centre for Medium-Range Weather Forecasts (ECMWF) 3 . Our method also works well with extreme weather forecasts and ensemble forecasts. When initialized with reanalysis data, the accuracy of tracking tropical cyclones is also higher than that of ECMWF-HRES.

Similar content being viewed by others

Deep learning for twelve hour precipitation forecasts

Neural general circulation models for weather and climate

A machine learning model that outperforms conventional global subseasonal forecast models

Weather forecasting is an important application of scientific computing that aims to predict future weather changes, especially in regards to extreme weather events. In the past decade, high-performance computing systems have greatly accelerated research in the field of numerical weather prediction (NWP) methods 1 . Conventional NWP methods are primarily concerned with describing the transitions between discretized grids of atmospheric states using partial differential equations (PDEs) and then solving them with numerical simulations 4 , 5 , 6 . These methods are often slow; a single simulation for a ten-day forecast can take hours of computation in a supercomputer that has hundreds of nodes 7 . In addition, conventional NWP algorithms largely rely on parameterization, which uses approximate functions to capture unresolved processes, where errors can be introduced by approximation 8 , 9 .

The rapid development of deep learning 10 has introduced a promising direction, which the scientific community refers to as artificial intelligence (AI)-based methods 2 , 11 , 12 , 13 , 14 , 15 , 16 . Here, the methodology is to train a deep neural network to capture the relationship between the input (reanalysis weather data at a given point in time) and the output (reanalysis weather data at the target point in time). On specialized computational devices such as graphics processing units (GPUs), AI-based methods are extremely fast. To give a recent example, FourCastNet 2 takes only 7 s to compute a 100-member, 24-hour forecast, which is orders of magnitudes faster than conventional NWP methods. However, the accuracy of FourCastNet is still below satisfactory; its root mean square error (RMSE) of a 5-day Z500 (500 hPa geopotential) forecast is 484.5, which is much worse than the 333.7 reported by the operational integrated forecasting system (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF) 3 . In a recent survey 17 , researchers agreed that AI holds great potential, but admitted that “a number of fundamental breakthroughs are needed” before AI-based methods can beat NWP.

These breakthroughs seem to be happening earlier than expected. Here we present Pangu-Weather (see Methods for an explanation of the name ‘Pangu’), a powerful AI-based weather forecasting system that produces stronger deterministic forecast results than the operational IFS on all tested weather variables against reanalysis data. Our technical contributions are two-fold. First, we integrated height information into a new dimension so that the input and output of our deep neural networks can be conceptualized in three dimensions. We further designed a three-dimensional (3D) Earth-specific transformer (3DEST) architecture to inject Earth-specific priors into the deep networks. Our experiments show that 3D models, by formulating height into an individual dimension, have the ability to capture the relationship between atmospheric states in different pressure levels and thus yield significant accuracy gains, compared with two-dimensional models such as FourCastNet 2 . Second, we applied a hierarchical temporal aggregation algorithm that involves training a series of models with increasing forecast lead times. Hence, in the testing stage, the number of iterations needed for medium-range weather forecasting was largely reduced, and the cumulative forecast errors were alleviated. Experiments on the fifth generation of ECMWF reanalysis (ERA5) data 18 validated that Pangu-Weather is good at deterministic forecast and extreme weather forecast while being more than 10,000-times faster than the operational IFS.

Global weather forecasting with 3D networks

We established our weather forecast system via deep learning. The methodology involves training deep neural networks to take reanalysis weather data at a given point in time as input, and then produce reanalysis weather data at a future point in time as output. We used a single point in time for both input and output. The time resolution of the ERA5 data is 1 h; in the training subset (1979–2017), there were as many as 341,880 time points, the amount of training data in one epoch. To alleviate the risk of over-fitting, we randomly permuted the order of sample from the training data at the start of each epoch. We trained four deep networks with lead times (the time difference between input and output) at 1 h, 3 h, 6 h and 24 h, respectively. Each of the four deep networks was trained for 100 epochs, and each of them takes approximately 16 days on a cluster of 192 NVIDIA Tesla-V100 GPUs.

The architecture of our deep network is shown in Fig. 1a . This architecture is known as the 3D Earth-specific transformer (3DEST). We fed all included weather variables, including 13 layers of upper-air variables and the surface variables, into a single deep network. We then performed patch embedding to reduce the spatial resolution and combined the down-sampled data into a 3D cube. The 3D data are propagated through an encoder–decoder architecture derived from the Swin transformer 19 , a variant of a vision transformer 20 , which has 16 blocks. The output is split into upper-air variables and surface variables and is up-sampled with patch recovery to restore the original resolution. To inject Earth-specific priors into the deep network, we designed an Earth-specific positional bias (a mechanism of encoding the position of each unit; detailed in Methods ) to replace the original relative positional bias of Swin. This modification increases the number of bias parameters by a factor of 527, with each 3D deep network containing approximately 64 million parameters. Compared with the baseline, however, 3DEST has the same computational cost and has a faster convergence speed.

a , 3DEST architecture. Based on the standard encoder–decoder design of vision transformers, we adjusted the shifted-window mechanism 19 and applied an Earth-specific positional bias. b , Hierarchical temporal aggregation. Once given a lead time, we used a greedy algorithm to perform forecasting with as few steps as possible. We use FM1, FM3, FM6 and FM24 to indicate the forecast models with lead times being 1 h, 3 h, 6 h or 24 h, respectively. A 0 is the input weather state and \({\hat{{\bf{A}}}}_{t}\) denotes the predicted weather state at time t (in hours).

The lead time of a medium-range weather forecast is 7 days or longer. This prompted us to call the base deep networks (lead times being 1 h, 3 h, 6 h or 24 h) iteratively, using each forecasted result as the input of the next step. To reduce the cumulative forecast errors, we introduced hierarchical temporal aggregation, a greedy algorithm that always calls for the deep network with the largest affordable lead time. Mathematically, this greatly reduces the number of iterations. As an example, when the lead time was 56 h, we would execute the 24-hour forecast model 2 times, the 6-hour forecast model 1 time, and the 1-hour forecast model 2 times (Fig. 1b ). Compared with FourCastNet 2 , which uses a fixed 6-hour forecast model, our method is faster and more accurate. The limitation of this strategy is discussed in Methods .

Experimental setting and main results

We evaluated Pangu-Weather on the ERA5 data 18 , which is considered the best known estimation for most atmospheric variables 21 , 22 . To fairly compare Pangu-Weather against FourCastNet 2 , we trained our 3D deep networks on 39 years of data (from 1979 to 2017), validated them on 2019 data and tested them on 2018 data. We studied 69 factors, including 5 upper-air variables at 13 pressure levels (50 hPa, 100 hPa, 150 hPa, 200 hPa, 250 hPa, 300 hPa, 400 hPa, 500 hPa, 600 hPa, 700 hPa, 850 hPa, 925 hPa and 1,000 hPa) and 4 surface variables. When tested against reanalysis data, for each tested variable, Pangu-Weather produces a lower RMSE and a higher anomaly correlation coefficient (ACC) than the operational IFS and FourCastNet, the best NWP and AI-based methods, respectively. In particular, with a single-member forecast, Pangu-Weather reports an RMSE of 296.7 for a 5-day Z500 forecast, which is lower than that for the operational IFS and FourCastNet, which reported 333.7 and 462.5, respectively. In addition, the inference cost of Pangu-Weather is 1.4 s on a single GPU, which is more than 10,000-times faster than the operational IFS and on par with FourCastNet. Pangu-Weather not only produces strong quantitative results (for example, RMSE and ACC) but also preserves sufficient details for investigating certain extreme weather events. To demonstrate this capability, we studied the important application of tropical cyclone tracking. By finding the local minimum of mean sea-level pressure (MSLP), one of the surface variables, our algorithm achieved high accuracy in tracking 88 named tropical cyclones in 2018, including some (for example, Typhoon Kong-rey and Typhoon Yutu) that remain a challenge for the world’s best tracking systems, such as ECMWF-HRES (where HRES stands for high-resolution). Our research sheds light on AI-based medium-range weather forecasting systems and advances the progress on the path towards establishing AI as a complement to or surrogate for NWP, an achievement that was previously thought to be far off in the future 17 .

Deterministic global weather forecast

We performed the deterministic forecasting on the unperturbed initial states from ERA5. We then compared Pangu-Weather to the strongest methods in both NWP and AI, namely the operational IFS of ECMWF (data downloaded from the TIGGE (THORPEX Interactive Grand Global Ensemble) archive 3 ) and FourCastNet 2 . The spatial resolution of Pangu-Weather, 0.25° × 0.25°, was determined by the training data, which is comparable to the control forecast of ECMWF ENS 5 and identical to FourCastNet. The spacing of the forecast (the minimal unit of forecast time) of Pangu-Weather is 1 h, 6 times less than FourCastNet.

The overall forecast results for 2018 are shown in Fig. 2 . For each tested variable, including upper-air and surface variables, Pangu-Weather reports more accurate results than both the operational IFS and FourCastNet (when the variable is reported). In terms of RMSE (lower is better), Pangu-Weather typically reports 10%-lower values than operational IFS and 30%-lower values than FourCastNet. The advantage persists across all lead times (from 1 h to 168 h, that is, 7 days), and for some variables such as Z500, the advantage becomes more significant with a greater lead time. For quantitative studies in the Northern Hemisphere, the Southern Hemisphere and the tropics, refer to the Extended Data Figs. 1 – 3 . For the forecast results for 2020 and 2021 and the comparison with the results for 2018, refer to Extended Data Fig. 4 .

Ten variables were compared in terms of latitude-weighted RMSE (lower is better) and ACC (higher is better), where the first five variables were reported in FourCastNet and the last five were not. Here, Z500, T500, Q500, U500 and V500 indicate the geopotential, temperature, specific humidity, and the u -component and v -component of wind speed at 500 hPa, respectively. Z850 and T850 indicate the geopotential and temperature at 850 hPa, respectively. T2M indicates the 2-m temperature, and U10 and V10 indicate the u -component and v -component of 10-m wind speed, respectively.

Source data

To demonstrate our advantage, we introduced a concept called ‘forecast time gain’, which corresponds to the average difference between the lead times of Pangu-Weather and a competitor when they report the same accuracy. Pangu-Weather typically shows a forecast time gain of 10–15 h over the operational IFS, and for some variables such as specific humidity, the gain is more than 24 h. This implies the difficulty that conventional NWP methods have when forecasting specific variables, yet AI-based methods benefit by learning effective patterns from an abundance of training data. Compared with FourCastNet, the forecast time gain of Pangu-Weather is as great as 40 h, showing the significant advantage of our technical design, resulting especially from the 3D deep networks and the advanced temporal aggregation strategy. The forecast time gains of Pangu-Weather in terms of different weather variables are summarized in Extended Data Table 2 .

Figure 3 shows a visualization of the 3-day forecast results of Pangu-Weather. We studied two upper-air variables, Z500 and T850 (850 hPa temperature), and two surface variables, 2-m temperature and 10-m wind speed, and compared the results with the operational IFS and the ERA5 ground truth. The results of both Pangu-Weather and operational IFS are sufficiently close to the ground truth, yet there are visible differences between them. Pangu-Weather produced smoother contour lines, implying that the model tends to forecast similar values for neighbouring regions. It is a general property of any regression algorithm (including deep neural networks) to converge on average values. In contrast, the operational IFS forecast is less smooth, because it calculates a single estimated value at each grid cell by solving a system of PDEs with initial conditions, while the chaotic nature of weather and the inevitably imprecise knowledge of the initial conditions and sub-grid scale processes can cause statistical uncertainties in each forecast.

The 3-day forecast of two upper-air variables (Z500 and T850) and two surface variables (2-m temperature and 10-m wind speed). For each case, Pangu-Weather (left), the operational IFS 3 (middle) and the ERA5 ground truth 18 (right) are shown. For all cases, the input time is 00:00 UTC on 1 September 2018.

Tracking tropical cyclones

Next, we used Pangu-Weather to track tropical cyclones. Given an initial time point, we set the lead time to be multiples of 6 h (ref. 23 ) and initiated Pangu-Weather to forecast future weather states. We looked for the local minimum of MSLP that satisfied certain conditions, such as the cyclone eye. The tracking algorithm is described in the supplementary material for this paper. We used the International Best Track Archive for Climate Stewardship (IBTrACS) project 24 , 25 , which contains the best available estimations for tropical cyclones.

We compared Pangu-Weather with ECMWF-HRES, a strong cyclone tracking method based on high-resolution (9 km × 9 km) operational weather forecasting. We chose 88 named tropical cyclones in 2018 that appear in both IBTrACS and ECMWF-HRES. As shown in Fig. 4 , Pangu-Weather statistically produced more accurate tracking results than ECMWF-HRES for these cyclones. The 3-day and 5-day mean direct position errors for cyclone eyes were reported at 120.29 km and 195.65 km for Pangu-Weather, which are smaller than 162.28 km and 272.10 km for ECMWF-HRES, respectively. The breakdowns of tracking errors with respect to regions and intensities are provided in Extended Data Fig. 5 . The advantage of Pangu-Weather becomes more significant as the lead time increases. We also show the tracking results of the two strongest cyclones in the western Pacific, Kong-rey and Yutu, in Fig. 4 . See the supplementary material for a detailed analysis.

a , b , Tracking results for two strong tropical cyclones in 2018, that is, Typhoon Kong-rey (2018–25) and Yutu (2018–26). The initial time point is shown below each panel. The time gap between neighbouring dots is 6 h. Pangu-Weather forecasts the correct path of Yutu (that is, it goes to the Philippines) at 12:00 UTC on 23 October 2018, whereas ECMWF-HRES obtains the same conclusion 2 days later, before which it predicts that Yutu will make a big turn to the northeast. c , A comparison between Pangu-Weather and ECMWF-HRES in terms of mean direct position error over 88 cyclones in 2018. Each number in brackets in the x -axis indicates the number of samples used to calculate the average. For example, ‘(788)’ means that there are in total 788 initial points from which the typhoon lasts for at least 24 hours, and the 788 direct position errors of Pangu-Weather and ECMWF-HRES were averaged into the final results. Panels a and b were plotted using the Matplotlib Basemap toolkit.

Despite the promising tracking results, we point that the direct comparison between Pangu-Weather and ECMWF-HRES is somewhat unfair, because ECMWF-HRES used the IFS initial condition data as its input, whereas Pangu-Weather used reanalysis data.

Ensemble weather forecast

As an AI-based method, Pangu-Weather is more than 10,000-times faster than the operational IFS. This offers an opportunity for performing large-member ensemble forecasts with small computational costs. We investigated FourCastNet 2 to study a preliminary ensemble method that adds perturbations to initial weather states. We then generated 99 random perturbations (detailed in Methods ) and added them to the unperturbed initial state. Thus, we obtained a 100-member ensemble forecast by simply averaging the forecast results. As shown in Fig. 5 , for each variable, the ensemble mean is slightly worse than the single-member method in the short-range (for example, 1 day) weather forecasts, but significantly better when the lead time is 5–7 days. This aligns with FourCastNet 2 , indicating that large-member ensemble forecasts are especially useful when single-model accuracy is lower, yet they present the risk of introducing unexpected noise to short-range forecasts. Ensemble forecasting presents more benefits to non-smooth variables such as Q500 (500 hPa specific humidity) and U10 (10 m u -component of wind speed). In addition, the spread-skill ratio of Pangu-Weather is smaller than 1, indicating that the current ensemble method is somewhat underdispersive. Compared with NWP methods, Pangu-Weather largely reduces the cost of ensemble forecasting, allowing meteorologists to apply their expertise to control noise and improve ensemble forecast accuracy.

The RMSE of the ensemble mean forecast (lower is better) for three upper-air variables (Z500, Q500 and U500) and two surface variables (T2M and U10). We also followed a recent work 35 to plot two metrics, the CRPS (lower is better) and the spread-skill ratio (an ideal ensemble model produces spread-skill ratios of 1.0, shown as the dashed lines), which further demonstrate the properties of our ensemble forecast results. Here, Z500, Q500 and U500 indicate the geopotential, temperature and the u -component of wind speed at 500 hPa, respectively. T2M indicates the 2-m temperature and U10 indicates the u -component of 10-m wind speed.

In this paper, we present Pangu-Weather, an AI-based system that trains deep networks for fast and accurate numerical weather forecasting. The major technical contributions include the design of the 3DEST architecture and the application of the hierarchical temporal aggregation strategy for medium-range forecasting. By training the models on 39 years of global weather data, Pangu-Weather produces better deterministic forecast results on reanalysis data than the world’s best NWP system, the operational IFS of ECMWF, while also being much faster. In addition, Pangu-Weather is excellent at forecasting extreme weather events and performing ensemble weather forecasts. Pangu-Weather reveals the potential of using large pre-trained models for various downstream applications, showing the same trend as other AI scopes, such as computer vision 26 , 27 , natural language processing 28 , 29 , cross-modal understanding 30 and beyond.

Despite the promising forecast accuracy on reanalysis data, our algorithm has some limitations. First, throughout this paper, Pangu-Weather was trained and tested on reanalysis data, but real-world forecast systems work on observational data. There are differences between these data sources; thus, Pangu-Weather’s performance across applications needs further investigation. Second, some weather variables, such as precipitation, were not investigated in this paper. Omitting these factors may cause the current model to lack some abilities, for example, making use of precipitation data for the accurate prediction of small-scale extreme weather events, such as tornado outbreaks 31 , 32 . Third, AI-based methods produce smoother forecast results, increasing the risk of underestimating the magnitudes of extreme weather events. We studied a special case, cyclone tracking, but there is much more work to do. Fourth, temporal inconsistency can be introduced by using models with different lead times. This is a challenging topic worth further investigation.

Looking to the future, there is room for improvement for both AI-based methods and NWP methods. On the AI side, further gains can be found by incorporating more vertical levels and/or atmospheric variables, integrating the time dimension and training four-dimensional deep networks 33 , 34 , using deeper and/or wider networks, or simply increasing the number of training epochs. All of these directions call for more powerful GPU clusters with larger memories and higher FLOPS (floating point operations per second), which is the current trend of the AI community. On the NWP side, post-processing methods can be developed to alleviate the predictable biases of NWP models. We expect that AI-based and NWP methods will be combined in the future to bring about even stronger performance.

Mathematical settings

We denoted all studied global weather variables at time t as A t . This is a 3D matrix of size N lon  ×  N lat  × 69, where N lon  = 1,440 and N lat  = 721 are the spatial resolution along the longitude and latitude axes, respectively, and 69 is the number of studied variables. In other words, each horizontal pixel occupies 0.25° × 0.25° on Earth’s surface. The mathematical problem is that given the forecast time point t 0 , assume that A t for all t  ≤  t 0 are available, the algorithm is asked to predict \({{\bf{A}}}_{{t}_{0}+\Delta t}\) where Δ t is called the lead time. Owing to the limitation of GPU memory, in our work, the forecast algorithm only used \({{\bf{A}}}_{{t}_{0}}\) as input and predicted \({{\bf{A}}}_{{t}_{0}+\Delta t}\) as output. For this purpose, we trained a deep neural network, \(f({{\bf{A}}}_{{t}_{0}}\,;{\boldsymbol{\theta }})\) , where θ denotes the learnable parameters.

Evaluation metrics

When the predicted version of A t is available ( t   =   t 0  + Δ t ), denoted as \({\hat{{\bf{A}}}}_{t}\) , we computed two metrics, RMSE and ACC, defined as follows:

Here, v is any weather variable, \({{\bf{A}}}_{i,j,t}^{v}\) is a scalar representing the value of v at time t and horizontal coordinate ( i ,  j ). \(L\left(i\right)={N}_{{\rm{lat}}}\times \frac{{\rm{\cos }}{\phi }_{i}}{{\sum }_{{i\text{'}}=1}^{{N}_{{\rm{lat}}}}{\rm{\cos }}{\phi }_{{i\text{'}}}}\) is the weight at latitude ϕ i . A ′ denotes the difference between A and the climatology, that is, the long-term mean of weather states that are estimated on the training data over 39 years. The RMSE and ACC values were averaged over all times and horizontal coordinates to produce the average numbers for variable v and lead time Δ t . The RMSE and ACC metrics can also be evaluated for specific regions, for example, in the Northern Hemisphere, the Southern Hemisphere and the tropics. Refer to Fig. 2 and Extended Data Figs. 1 – 3 for the overall and breakdown results in 2018.

Ensemble forecast metrics

We follow a recent work 35 to compute two metrics for ensemble weather forecast, namely, the continuous ranked probability score (CRPS) and the spread-skill ratio (SSR). Mathematically, CRPS is defined as

where F (·)denotes the cumulative distribution function of the forecast distribution and \({\mathbb{I}}\) (·) is an indicator function that takes a value of 1 if the statement is true and 0 otherwise. We follow the original paper and use the xskillscore Python package for computing the CRPS. SSR is obtained by dividing ‘spread’ by RMSE with spread being

Here, var(·) indicates the variance in the ensemble dimension. The spread and RMSE values averaged over all forecasts are used for computing SSR. If an ensemble is perfectly reliable, it should report an SSR of 1.0.

Data preparation details

The ERA5 dataset 18 contains global, hourly reanalysis data for the past 60 years. The observation data and the prediction of numerical models are blended into reanalysis data using numerical assimilation methods, providing a high-quality benchmark for global weather forecasting. We made use of the reanalysis data of every single hour so that the algorithm can perform hourly weather forecasting. We kept the highest spatial resolution available in ERA5, 0.25° × 0.25° on Earth’s sphere, resulting in an input resolution of 1,440 × 721: the latitude dimension has an extra entry because the northernmost and southernmost positions do not overlap.

We followed WeatherBench 13 to choose 13 out of 37 pressure levels (50 hPa, 100 hPa, 150 hPa, 200 hPa, 250 hPa, 300 hPa, 400 hPa, 500 hPa, 600 hPa, 700 hPa, 850 hPa, 925 hPa and 1,000 hPa) and the surface level. To fairly compare with the online version of the ECMWF control forecast, we chose to forecast the factors published in the TIGGE dataset 3 , namely, five upper-air variables (geopotential, specific humidity, temperature, and the u -component and v -component of wind speed) and four surface variables (2-m temperature, the u -component and v -component of 10-m wind speed, and MSLP). For a complete list of studied variables and the corresponding abbreviations, refer to Extended Data Table 1 . In addition, three constant masks (the topography mask, land–sea mask and soil-type mask) were added to the input of surface variables.

When we prepared for the test data in 2018, we excluded the test points on 1 January 2018 owing to the overlap with training data. In addition, all test points in December 2018 are unavailable for the upper-air variables owing to a server error of ECMWF. FourCastNet also excluded these data from the testing phase.

Deep network details

There are two sources of input and output data, namely, upper-air variables and surface variables. The former involves 13 pressure levels, each of which has 5 variables, and they together form a 13 × 1,440 × 721 × 5 volume. The latter contains a 1,440 × 721 × 4 volume. These parameters were first embedded from the original space into a C -dimensional latent space. We used a common technique named patch embedding for dimensionality reduction. For the upper-air part, the patch size is 2 × 4 × 4 so the embedded data have a shape of 7 × 360 × 181 ×  C . For the surface variables, the patch size is 4 × 4 so the embedded data have a shape of 360 × 181 ×  C , where C is the base channel width and was set to be 192 in our work. These two data volumes were then concatenated along the first dimension to yield a 8 × 360 × 181 ×  C volume. The volume was then propagated through a standard encoder–decoder architecture with 8 encoder layers and 8 decoder layers. The output of the decoder is still a 8 × 360 × 181 ×  C volume, which was projected back to the original space with patch recovery, producing the desired output. Below, we describe the technical details of each component.

Patch embedding and patch recovery

We followed the standard vision transformer to use a linear layer with  GELU (Gaussian Error Linear Unit) activation for patch embedding. In our implementation, a patch has 2 × 4 × 4 pixels for upper-air variables and 4 × 4 for surface variables. The stride of sliding windows is the same as the patch size, and necessary zero-value padding was added when the data size is indivisible by the patch size. The number of parameters for patch embedding is (4 × 4 × 2 × 5) ×  C for upper-air variables and (4 × 4 × 4) ×  C for surface variables. Patch recovery performs the opposite operation, having the same number of parameters but these parameters are not shared with patch embedding.

The encoder–decoder architecture

The data size remains unchanged as 8 × 360 × 181 ×  C for the first 2 encoder layers, whereas for the next 6 layers, the horizontal dimensions were reduced by a factor of 2 and the number of channels was doubled, resulting in a data size of 8 × 180 × 91 × 2 C . The decoder part is symmetric to the encoder part, with the first 6 decoder layers having a size of 8 × 180 × 91 × 2 C and the next 2 layers having a size of 8 × 360 × 181 ×  C . The outputs of the second encoder layer and the seventh decoder layer were concatenated along the channel dimension. We follow the implementation of Swin transformers 19 to connect the adjacent layers of different resolutions with down-sampling and up-sampling operations. For down-sampling, we merged four tokens into one (the feature dimensionality increases from C to 4 C ) and performed a linear layer to reduce the dimensionality to 2 C . For up-sampling, the reverse operations were performed.

3D Earth-specific transformer

Each encoder and decoder layer is a 3DEST block. It is similar to the standard vision transformer block 20 but specifically designed to align with Earth’s geometry. We used the standard self-attention mechanism of vision transformers. To further reduce computational costs, we inherited the window-attention mechanism 19 to partition the feature maps into windows, each of which contains at most 2 × 12 × 6 tokens. The shifted-window mechanism 19 was applied so that for every layer, the grid partition differs from the previous one by half window size. As coordinates in longitude direction are periodic, the half windows at the left and right edges are merged into one full window. The merge operation was not performed along the latitude direction because it is not periodic. We refer the reader to the original papers 19 , 20 for more details about vision transformers.

Earth-specific positional bias

Swin transformer 19 used a relative positional bias to represent the translation-invariant component of attentions, where the bias was computed upon the relative coordinate of each window. For global weather forecasting, however, the situation is a bit different: each token corresponds to an absolute position on Earth’s coordinate system; as the map is a projection of Earth’s sphere, the spacing between neighbouring tokens can be different. More importantly, some weather states are closely related to the absolute position. Examples of geopotential, wind speed and temperature are shown in Extended Data Fig. 6 . To capture these properties, we introduced an Earth-specific positional bias, which works by adding a positional bias to each token based on its absolute (rather than relative) coordinate.

Mathematically, let the entire feature map be a volume with a spatial resolution of N pl  ×  N lon  ×  N lat , where N pl , N lon and N lat indicate the size along the axes of pressure levels, longitude and latitude, respectively. The data volume was partitioned into M pl  ×  M lon  ×  M lat windows, and each window has a size of W pl  ×  W lon  ×  W lat . The Earth-specific positional bias matrix contains M pl  ×  M lat submatrices ( M lon does not appear here because different longitudes share the same bias: the longitude indices are cyclic and spacing is evenly distributed along this axis), each of which has \({W}_{{\rm{pl}}}^{2}\times {\left(2{W}_{{\rm{lon}}}-1\right)\times W}_{{\rm{lat}}}^{2}\) learnable parameters. When the attention was computed between two units within the same window, we used the indices of the pressure level and latitude, ( m pl ,  m lat ), to locate the corresponding bias submatrix. Then, we used the intra-window coordinates, \(\left({h}_{1}^{{\prime} },{\lambda }_{1}^{{\prime} },{\phi }_{1}^{{\prime} }\right)\) and \(\left({h}_{2}^{{\prime} },{\lambda }_{2}^{{\prime} },{\phi }_{2}^{{\prime} }\right)\) , to look up the bias value at \(\left({h}_{1}^{{\prime} }\,+\,{h}_{2}^{{\prime} }\times {W}_{{\rm{pl}}},{\lambda }_{1}^{{\prime} }-{\lambda }_{2}^{{\prime} }\,+\,{W}_{{\rm{lon}}}-1,{\phi }_{1}^{{\prime} }\,+\,{\phi }_{2}^{{\prime} }\,\times \,{W}_{{\rm{lat}}}\right)\) of the ( m pl ,  m lat )th submatrix.

Design choices

We briefly discuss other design choices. Owing to the large training overhead, we did not perform exhaustive studies on the hyperparameters and we believe that there exist configurations or hyperparameters that lead to higher accuracy. First, we used 8 (2 + 6) encoder and decoder layers, which is significantly fewer than the standard Swin transformer 19 . This is to reduce the complexity of both time and memory. If one has a more powerful cluster with larger GPU memory, increasing the network depth can bring higher accuracy. Second, it is possible to reduce the number of parameters used in the Earth-specific positional bias by parameter sharing or other techniques. However, we did not consider it a key issue, because it is unlikely to deploy the weather forecasting model to edge devices with limited storage. Third, it is possible and promising to feed the weather states of more time indices into the model, which changes all tensors from three dimensions to four dimensions. Although the AI community has shown the effectiveness of four-dimensional deep networks 33 , 34 , the limited available computational budget prevented us from exploring this method.

Optimization details

The four individual models were trained for 100 epochs using the Adam optimizer. We used the mean-absolute-error loss. The normalization was performed on each two-dimensional input field (for example, Z500) separately. It worked by subtracting the mean value from the two-dimensional field followed by dividing it by the standard deviation. The mean and standard deviation of each variable were computed on the weather data from 1979 to 2017. The weight for each variable was inversely proportional to the average loss value computed in an early run, which was designed to facilitate equivalence of the contributions by these variables. Specifically, the weights for upper-air variables were 3.00, 0.60, 1.50, 0.77 and 0.54 for Z, Q, T, U and V, respectively, and the weights for surface variables were 1.50, 0.77, 0.66 and 3.00 for MSLP, U10, V10 and T2M, respectively. We added a weight of 1.0 to the mean-absolute-error loss of the upper-air variables and 0.25 to that of the surface variables, and summed up the two losses. We used a batch size of 192 (that is, 1 training sample per GPU). The learning rate started with 0.0005 and gradually annealed to 0 following the cosine schedule. All starting time points in the training subset (1979–2017) were randomly permuted in each epoch to alleviate over-fitting. A weight decay of 3 × 10 −6 and ScheduledDropPath 36 with a drop ratio of 0.2 were adopted to alleviate over-fitting. We found that all models have not yet arrived at full convergence at the end of 100 epochs, so we expect that extending the training procedure can improve the forecast accuracy. We plotted the accuracy of some tested variables with respect to different lead times (1 h, 3 h, 6 h and 24 h) in Extended Data Fig. 7 .

Inference speed

The inference speed of Pangu-Weather is comparable to that of FourCastNet 2 . In a system-level comparison, FourCastNet requires 0.28 s for inferring a 24-hour forecast on a Tesla-A100 GPU (312 teraFLOPS), whereas Pangu-Weather needs 1.4 s on a Tesla-V100 GPU (120 teraFLOPS). Taking GPU performance into consideration, Pangu-Weather is about 50% slower than FourCastNet. Pangu-Weather is more than 10,000-times faster than the operational IFS, which requires several hours in a supercomputer with hundreds of nodes.

Computation of relative quantile error

We followed a previous work 37 to compare the values of top-level quantiles calculated on the forecast result and ground truth. Mathematically, we set D  = 50 percentiles, denoted as q 1 ,  q 2 , ...,  q D . We followed FourCastNet 2 to set q 1  = 90% and q D  = 99.99%, and the intermediate percentile values were linearly distributed between q 1 and q D in the logarithmic scale. Then, the corresponding quantiles, denoted as Q 1 ,  Q 2 , ...,  Q D , were computed individually for each pair of weather variable and lead time. For example, for all 3-day forecasts of the U10 variable, pixel-wise values were gathered from all frames for statistics. We followed FourCastNet 2 to plot the extreme percentiles with respect to lead time in Extended Data Fig. 7 .

Finally, the relative quantile error (RQE) was computed for measuring the overall difference between the ground truth and any weather forecast algorithm:

where Q d and \({\hat{Q}}_{d}\) are the d th quantile calculated on the ERA5 ground truth and the forecast algorithm being investigated. RQE can measure the overall tendency, where RQE < 0 and RQE > 0 imply that the forecast algorithm tends to underestimate and overestimate the intensity of extremes, respectively. We found that both Pangu-Weather and the operational IFS tend to underestimate extremes. Pangu-Weather suffers heavier underestimation as the lead time increases. It is noted that RQE and the individual quantile values have limitations: they do not evaluate whether extreme values occur at the right location and time, but only look at the value distribution. The ability of Pangu-Weather to capture individual extreme events was further validated with the experiments of tracking tropical cyclones.

Algorithm for tracking tropical cyclones

We followed a classical algorithm 38 that locates the local minimum of MSLP to track the eye of tropical cyclones. Given the starting time point and the corresponding initial position of a cyclone eye, we iteratively called for the 6-hour forecast algorithm and looked for a local minimum of MSLP that satisfies the following conditions:

There is a maximum of 850 hPa relative vorticity that is larger than 5 × 10 −5 within a radius of 278 km for the Northern Hemisphere, or a minimum that is smaller than −5 × 10 −5 for the Southern Hemisphere.

There is a maximum of thickness between 850 hPa and 200 hPa within a radius of 278 km when the cyclone is extratropical.

The maximum 10-m wind speed is larger than 8 m s −1 within a radius of 278 km when the cyclone is on land.

Once the cyclone’s eye is located, the tracking algorithm continued to find the next position in a vicinity of 445 km. The tracking algorithm terminated when no local minimum of MSLP is found to satisfy the above conditions. See Extended Data Fig. 8 for two tracking examples.

Tracking results in different subsets

We extended Fig. 4c by plotting the mean direct position errors with respect to different basins or different intensities in Extended Data Fig. 5 . In each subset, Pangu-Weather reports lower errors and the advantage becomes more significant with a greater lead time, aligning with the conclusions we drew from the entire dataset. Again, we emphasize that the comparison against ECMWF-HRES is somewhat unfair, because ECMWF-HRES used IFS initial condition data, whereas Pangu-Weather used reanalysis data.

More tropical cyclones

Below is a more detailed analysis of four tropical cyclones. The advantage of Pangu-Weather mainly lies in tracking cyclone paths in the early stages.

Typhoon Kong-rey (2018-25) is one of the most powerful tropical cyclones worldwide in 2018. As shown in Fig. 4 , ECMWF-HRES forecasts that Kong-rey would land in China, but it actually did not. Pangu-Weather, instead, produces accurate tracking results which almost coincide with the ground truth. Also, Extended Data Fig. 8 shows the tracking results of Pangu-Weather and ECMWF-HRES at different time points: the forecast result of Pangu-Weather barely changes with time, and ECMWF-HRES arrives at the conclusion that Kong-rey would not land in China more than 48 h later than Pangu-Weather.

Typhoon Yutu (2018-26) is an extremely powerful tropical cyclone that caused catastrophic destruction in the Mariana Islands and the Philippines. It ties with Kong-rey as the most powerful tropical cyclone worldwide in 2018. As shown in Fig. 4 , Pangu-Weather makes the correct forecast result (Yutu goes to the Philippines) as early as 6 days before landing, whereas ECMWF-HRES incorrectly predicts that Yutu will make a big turn to the northeast in the early stage. ECMWF-HRES produces the correct tracking results more than 48 h later than Pangu-Weather.

Hurricane Michael (2018-13) is the strongest hurricane of the 2018 Atlantic hurricane season. As shown in Extended Data Fig. 8 , with a starting time that is more than 3 days earlier than landing, both Pangu-Weather and ECMWF-HRES forecast the landfall in Florida. But, the delay of predicted landing time is only 3 h for Pangu-Weather whereas it is 18 h for ECMWF-HRES. In addition, Pangu-Weather shows great advantages in tracking Michael after it landed, whereas the tracking of ECMWF-HRES is shorter and obviously shifts to the east.

Typhoon Ma-on (2022-09) is a severe tropical storm that impacted the Philippines and China. As shown in Extended Data Fig. 8 , when the starting time point is about 3 days earlier than the landing, ECMWF-HRES produces a wrong forecast result that Ma-on would land in Zhuhai, China, whereas the forecast result of Pangu-Weather is close to the truth.

The better tracking results of Pangu-Weather are mainly inherited from the accurate deterministic forecast accuracy on reanalysis data. In Extended Data Fig. 8 , we show how Pangu-Weather tracks Hurricane Michael and Typhoon Ma-on following the specified tracking algorithm. Among the four variables, MSLP and 10-m wind speed were directly produced by deterministic forecast, and thickness and vorticity were derived from geopotential and wind speed. This indicates that Pangu-Weather can produce intermediate results that support cyclone tracking, which further assists meteorologists in understanding and exploiting the tracking results.

Random perturbations

Each perturbation generated for ensemble weather forecast contains 3 octaves of Perlin noise, with the scales being 0.2, 0.1 and 0.05, and the number of periods to generate along each axis (the longitude or the latitude) being 12, 24 and 48, respectively. We used the code provided in a GitHub repository ( https://github.com/pvigier/perlin-numpy ) and modified the code for acceleration. We added a section to the pseudocode.

Previous work

There are mainly two lines of research for weather forecasting. Throughout this paper, we have been using ‘conventional NWP’ or simply ‘NWP’ methods to refer to the numerical simulation methods, and use ‘AI-based’ methods to specify data-driven forecasting systems. We understand that, verbally, AI-based methods also belong to NWP, but we followed the convention 17 to use these terminologies.

NWP methods often partition the atmospheric states into discretized grids, use PDEs to describe the transition between them 1 , 39 , 40 and solve the PDEs using numerical simulations. The spacing of grids is key to forecast accuracy, but it is constrained by the computational budget and thus the spatial resolution of weather forecasts is often limited. Parameterization 41 is an effective method for capturing unresolved processes. NWP methods have been widely applied, but they are troubled by the super-linearly increasing computational overhead 1 , 42 and it is often difficult to perform efficient parallelization for them 43 . The heavy computational overhead of NWP also restricts the number of ensemble members, hence weakening the diversity and accuracy of probabilistic weather forecasts.

AI-based methods offer a complementary path for weather forecasting. The cutting-edge technology of AI lies in deep learning 10 , which assumes that the complex relationship between input and output data can be learned from abundant training data without knowing the actual physical procedure and/or formulae. In the scope of weather forecasting, AI-based methods were first applied to the problems of precipitation forecasting based on radar data 44 , 45 , 46 , 47 or satellite data 48 , 49 , where the traditional methods that are much influenced by the initial conditions were replaced by deep-learning-based methods. The powerful expressive ability of deep neural networks led to success in these problems, which further encouraged researchers to delve into medium-range weather forecasting 2 , 11 , 12 , 13 , 14 , 15 , 16 as a faster complement or surrogate of NWP methods. State-of-the-art deep-learning methods mostly rely on large models (that is, with large numbers of learnable parameters) to learn complex patterns from the training data.

The name of ‘Pangu’

Pangu is a primordial being and creation figure in Chinese mythology who separated heaven and earth and became geographic features such as mountains and rivers (see https://en.wikipedia.org/wiki/Pangu ). Pangu is also a series of pre-trained AI models developed by Huawei Cloud that covers computer vision, natural language processing, multimodal understanding, scientific computing (including weather forecasting) and so on.

Data availability

For training and testing Pangu-Weather, we downloaded a subset of the ERA5 dataset (around 60 TB) from https://cds.climate.copernicus.eu/ , the official website of Copernicus Climate Data (CDS). For comparison with operational IFS, we downloaded the forecast data and tropical cyclone tracking results of ECMWF from https://confluence.ecmwf.int/display/TIGGE , the official website of the TIGGE archive. We downloaded the ground-truth routes of tropical cyclones from the International Best Track Archive for Climate Stewardship (IBTrACS) project, https://www.ncei.noaa.gov/products/international-best-track-archive . All these data are publicly available for research purposes.  Source data are provided with this paper.

Code availability

The code base of Pangu-Weather was established on PyTorch, a Python-based library for deep learning. In building and optimizing the backbones, we made use of the code base of Swin transformer, available at https://github.com/microsoft/Swin-Transformer . Other details, including network architectures, modules, optimization tricks and hyperparameters, are available in the paper and the pseudocode. The computation of the CRPS metric relied on the xskillscore Python package, https://github.com/xarray-contrib/xskillscore/ . The implementation of Perlin noise was inherited from a GitHub repository, https://github.com/pvigier/perlin-numpy . We also used other Python libraries, such as NumPy and Matplotlib, in the research project. We released the trained models, inference code and the pseudocode of details to the public at a GitHub repository: https://github.com/198808xc/Pangu-Weather ( https://doi.org/10.5281/zenodo.7678849 ). The trained models allow the researchers to explore Pangu-Weather’s ability on either ERA5 initial fields or ECMWF initial fields, where the latter is more practical as it can be used as an API for almost real-time weather forecasting.

Change history

14 september 2023.

A Correction to this paper has been published: https://doi.org/10.1038/s41586-023-06545-z

Bauer, P., Thorpe, A. & Brunet, G. The quiet revolution of numerical weather prediction. Nature 525 , 47–55 (2015).

Article   ADS   CAS   PubMed   Google Scholar  

Pathak, J. et al. FourCastNet: a global data-driven high-resolution weather model using adaptive Fourier neural operators. Preprint at https://arxiv.org/abs/2202.11214 (2022).

Bougeault, P. et al. The THORPEX interactive grand global ensemble. Bull. Am. Meteorol. Soc. 91 , 1059–1072 (2010).

Article   ADS   Google Scholar  

Skamarock, W. C. et al. A Description of the Advanced Research WRF Version 2 (National Center For Atmospheric Research Mesoscale and Microscale Meteorology Division, 2005).

Molteni, F., Buizza, R., Palmer, T. N. & Petroliagis, T. The ECMWF ensemble prediction system: methodology and validation. Q. J. R. Meteorol. Soc. 122 , 73–119 (1996).

Ritchie, H. et al. Implementation of the semi-Lagrangian method in a high-resolution version of the ECMWF forecast model. Mon. Weather Rev. 123 , 489–514 (1995).

Bauer, P. et al. The ECMWF Scalability Programme: Progress and Plans (European Centre for Medium Range Weather Forecasts, 2020).

Allen, M. R., Kettleborough, J. A. & Stainforth, D. A. Model error in weather and climate forecasting. In ECMWF Predictability of Weather and Climate Seminar 279–304 (European Centre for Medium Range Weather Forecasts, 2022); http://www.ecmwf.int/publications/library/do/references/list/209 .

Palmer, T. N. et al. Representing model uncertainty in weather and climate prediction. Annu. Rev. Earth Planet. Sci. 33 , 163–193 (2005).

Article   ADS   MathSciNet   CAS   MATH   Google Scholar  

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521 , 436–444 (2015).

Weyn, J. A., Durran, D. R. & Caruana, R. Can machines learn to predict weather? Using deep learning to predict gridded 500‐hPa geopotential height from historical weather data. J. Adv. Model. Earth Syst. 11 , 2680–2693 (2019).

Scher, S. & Messori, G. Weather and climate forecasting with neural networks: using general circulation models (GCMs) with different complexity as a study ground. Geosci. Model Dev. 12 , 2797–2809 (2019).

Rasp, S. et al. WeatherBench: a benchmark data set for data‐driven weather forecasting. J. Adv. Model. Earth Syst. 12 , e2020MS002203 (2020).

Weyn, J. A., Durran, D. R., Caruana, R. & Cresswell‐Clay, N. Sub‐seasonal forecasting with a large ensemble of deep‐learning weather prediction models. J. Adv. Model. Earth Syst. 13 , e2021MS002502 (2021).

Keisler, R. Forecasting global weather with graph neural networks. Preprint at https://arxiv.org/abs/2202.07575 (2022).

Hu, Y., Chen, L., Wang, Z. & Li, H. SwinVRNN: a data-driven ensemble forecasting model via learned distribution perturbation. J. Adv. Model. Earth Syst . 15 , e2022MS003211(2023).

Schultz, M. G. et al. Can deep learning beat numerical weather prediction? Phil. Trans. R. Soc. A 379 , 20200097 (2021).

Article   ADS   MathSciNet   CAS   PubMed   PubMed Central   Google Scholar  

Hersbach, H. et al. The ERA5 global reanalysis. Q. J. R. Meteorol. Soc. 146 , 1999–2049 (2020).

Liu, Z. et al. Swin transformer: hierarchical vision transformer using shifted windows. In Proc. International Conference on Computer Vision 10012–10022 (IEEE, 2021).

Dosovitskiy, A. et al. An image is worth 16x16 words: transformers for image recognition at scale. Preprint at https://arxiv.org/abs/2010.11929 (2020).

Betts, A. K., Chan, D. Z. & Desjardins, R. L. Near-surface biases in ERA5 over the Canadian Prairies. Front. Environ. Sci. 7 , 129 (2019).

Article   Google Scholar  

Jiang, Q. et al. Evaluation of the ERA5 reanalysis precipitation dataset over Chinese mainland. J. Hydrol. 595 , 125660 (2021).

Magnusson, L. et al. Tropical Cyclone Activities at ECMWF (European Centre for Medium Range Weather Forecasts, 2021).

Knapp, K. R., Kruk, M. C., Levinson, D. H., Diamond, H. J. & Neumann, C. J. The international best track archive for climate stewardship (IBTrACS) unifying tropical cyclone data. Bull. Am. Meteorol. Soc. 91 , 363–376 (2010).

Knapp, K. R., Diamond, H. J., Kossin, J. P., Kruk, M. C. & Schreck, C. J. International Best Track Archive for Climate Stewardship (IBTrACS) Project, Version 4 (NOAA National Centers for Environmental Information, 2018).

He, K., Fan, H., Wu, Y., Xie, S. & Girshick, R. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 9729–9738 (IEEE, 2020).

Bao, H., Dong, L. & Wei, F. Beit: BERT pre-training of image transformers. Preprint at https://arxiv.org/abs/2106.08254 (2021).

Devlin, J., Chang, M. W., Lee, K. & Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. Conference North American Chapter of the Association of Computational Linguistics Vol. 1, 4171–4186 (NAACL, 2019).

Brown, T. et al. Language models are few-shot learners. Adv. Neural. Inf. Process. Syst . 33 , 1877–1901 (2020).

Radford, A. et al. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning 8748–8763 (PMLR, 2021).

Chasteen, M. B. & Koch, S. E. Multiscale aspects of the 26–27 April 2011 tornado outbreak. Part I: outbreak chronology and environmental evolution. Mon. Weather Rev. 150 , 309–335 (2022).

Chasteen, M. B. & Koch, S. E. Multiscale aspects of the 26–27 April 2011 tornado outbreak. Part II: environmental modifications and upscale feedbacks arising from latent processes. Mon. Weather Rev. 150 , 337–368 (2022).

Choy, C., Gwak, J. Y. & Savarese, S. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 3075–3084 (IEEE, 2019).

Zhang S., Guo, S., Huang, W., Scott M. R. & Wang, L. V4D: 4D convolutional neural networks for video-level representation learning. Preprint at https://arxiv.org/abs/2002.07442 (2020).

Garg, S., Rasp, S. & Thuerey, N. WeatherBench probability: a benchmark dataset for probabilistic medium-range weather forecasting along with deep learning baseline models. Preprint at https://arxiv.org/abs/2205.00865 (2022).

Zoph, B., Vasudevan, V., Shlens, J. & Le, Q. V. Learning transferable architectures for scalable image recognition. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 8697–8710 (IEEE, 2017).

Fildier, B., Collins, W. D. & Muller, C. Distortions of the rain distribution with warming, with and without self‐aggregation. J. Adv. Model. Earth Syst. 13 , e2020MS002256 (2021).

Article   ADS   PubMed   PubMed Central   Google Scholar  

White, P. Newsletter No. 102—Winter 2004/05 (European Centre for Medium Range Weather Forecasts, 2005); https://www.ecmwf.int/node/14623 (2005).

Kalnay, E. Atmospheric Modeling, Data Assimilation and Predictability (Cambridge Univ. Press, 2003).

Lynch, P. The origins of computer weather prediction and climate modeling. J. Comput. Phys. 227 , 3431–3444 (2009).

Article   ADS   MathSciNet   MATH   Google Scholar  

Stensrud, D. J. Parameterization Schemes: Keys to Understanding Numerical Weather Prediction Models (Cambridge Univ. Press 2009).

Nakaegawa, T. High-performance computing in meteorology under a context of an era of graphical processing units. Computers 11 , 114 (2022).

Shi, X. et al. Convolutional LSTM network: a machine learning approach for precipitation nowcasting. Adv. Neural. Inf. Process. Syst. 28 , 802–810 (2015).

Google Scholar  

Shi, X. et al. Deep learning for precipitation nowcasting: a benchmark and a new model. Adv. Neural. Inf. Process. Syst. 30 , 5617–5627 (2017).

Agrawal, S. et al. Machine learning for precipitation nowcasting from radar images. Preprint at https://arxiv.org/abs/1912.12132 (2019).

Ravuri, S. et al. Skilful precipitation nowcasting using deep generative models of radar. Nature 597 , 672–677 (2021).

Article   ADS   CAS   PubMed   PubMed Central   Google Scholar  

Lebedev, V. et al. Precipitation nowcasting with satellite imagery. In Proc. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 2680–2688 (ACM, 2019).

Sønderby, C. K. et al. Metnet: a neural weather model for precipitation forecasting. Preprint at https://arxiv.org/abs/2003.12140 (2020).

Download references

Acknowledgements

We thank ECMWF for offering the ERA5 dataset and the TIGGE archive; NOAA National Centers for Environmental Information for the IBTrACS dataset; and other members of the Pangu team for discussions and support with the GPUs. Our appreciation also goes to the Integration Verification team of Huawei Cloud EI, which offers us a platform of high-performance parallel computing.

Author information

Authors and affiliations.

Huawei Cloud, Shenzhen, China

Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu & Qi Tian

You can also search for this author in PubMed   Google Scholar

Contributions

K.B. designed the project and trained the 3D deep networks for Pangu-Weather. L.X. improved the technical design. H.Z., X.C. and X.G. established the test environment and prepared for data. Q.T. managed and oversaw the research project. K.B. and L.X. wrote the paper.

Corresponding author

Correspondence to Qi Tian .

Ethics declarations

Competing interests.

K.B., L.X., H.Z., X.C., X.G. and Q.T. are employees of Huawei Cloud. A provisional patent (not granted an ID yet) was filed covering the generative algorithm described in this paper, listing the authors K.B., L.X. and Q.T. as inventors.

Peer review

Peer review information.

Nature thanks Matthew Chantry, Imme Ebert-Uphoff and Martin Schultz for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended data fig. 1 deterministic forecast results in the northern hemisphere..

We only compared Pangu-Weather to operational IFS 3 because FourCastNet 2 did not report the breakdown results. We followed ECMWF to define the “Northern Hemisphere” to be the region between latitude of 20° (exclusive) and 90° (inclusive). Here, Z500/T500/Q500/U500/V500 indicates the geopotential, temperature, specific humidity, and u -component and v -component of wind speed at 500 hPa. Z850/T850 indicates the geopotential and temperature at 850 hPa. T2M indicates the 2 m temperature, and U10/V10 indicates the u -component and v -component of 10 m wind speed.

Extended Data Fig. 2 Deterministic forecast results in the Southern Hemisphere.

We only compared Pangu-Weather to operational IFS 3 because FourCastNet 2 did not report the breakdown results. We followed ECMWF to define the “Northern Hemisphere” to be the region between latitude of −20° (exclusive) and −90° (inclusive). Here, Z500/T500/Q500/U500/V500 indicates the geopotential, temperature, specific humidity, and u -component and v -component of wind speed at 500 hPa. Z850/T850 indicates the geopotential and temperature at 850 hPa. T2M indicates the 2 m temperature, and U10/V10 indicates the u -component and v -component of 10 m wind speed.

Extended Data Fig. 3 Deterministic forecast results in the tropics.

We only compared Pangu-Weather to operational IFS 3 because FourCastNet 2 did not report the breakdown results. We followed ECMWF to define the “tropics” to be the region between latitude of +20° (inclusive) and −20° (inclusive). Here, Z500/T500/Q500/U500/V500 indicates the geopotential, temperature, specific humidity, and u -component and v -component of wind speed at 500 hPa. Z850/T850 indicates the geopotential and temperature at 850 hPa. T2M indicates the 2 m temperature, and U10/V10 indicates the u -component and v -component of 10 m wind speed.

Extended Data Fig. 4 Deterministic forecast results of Pangu-Weather in 2018, 2020 and 2021.

The RMSE and ACC values and trends are close among the three years, indicating Pangu-Weather’s stable forecasting skill over different years. Here, Z500/T500/Q500/U500/V500 indicates the geopotential, temperature, specific humidity, and u -component and v -component of wind speed at 500 hPa. Z850/T850 indicates the geopotential and temperature at 850 hPa. T2M indicates the 2 m temperature, and U10/V10 indicates the u -component and v -component of 10m wind speed.

Extended Data Fig. 5 Breakdowns of the mean direct position errors of tracking tropical cyclones.

a ) The breakdown into six oceans. b ) The breakdown into three intensity intervals. The overall statistics is displayed in Fig. 4c .

Extended Data Fig. 6 The motivation of using an Earth-specific positional bias.

a ) The horizontal map corresponds to an uneven spatial distribution on Earth’s sphere. b ) The geopotential height is closely related to the latitude. c ) The mean wind speed and temperature are closely related to the height (formulated as pressure levels). Sub-figures b ) and c ) were plotted using statistics on the ERA5 data.

Extended Data Fig. 7 Properties of deterministic forecast results.

a ) Single-model test errors. It shows the test errors (in RMSE) with respect to forecast time using single models (i.e., lead times being 1 h, 3 h, 6 h, and 24 h, respectively). Mind the accumulation of forecast errors as forecast time increases. b ) Visualization of the trend of quantiles with respect to lead time. It shows the trend of all the variables displayed in Fig. 2 and the comparisons to operational IFS 3 and ERA5 18 . Pangu-Weather often reports lower quantile values because AI-based methods tend to produce smooth forecasts. Here, Z500/T500/Q500/U500/V500 indicates the geopotential, temperature, specific humidity, and u -component and v -component of wind speed at 500 hPa. Z850/T850 indicates the geopotential and temperature at 850 hPa. T2M indicates the 2 m temperature, and U10/V10 indicates the u -component and v -component of 10 m wind speed.

Extended Data Fig. 8 Visualization of tracking tropical cyclones.

a ) The tracking results of cyclone eyes for Hurricane Michael (2018–13) and Typhoon Ma-on (2022–09) by Pangu-Weather and ECMWF-HRES, with a comparison to the ground-truth (by IBTrACS 24 , 25 ). b ) An illustration of the tracking process, where we used Pangu-Weather as an example. The algorithm locates the cyclone eye by checking four variables (from forecast results), namely, mean sea level pressure, 10 m wind speed, the thickness between 850 hPa and 200 hPa, and the vorticity of 850 hPa). The displayed figures correspond to the forecast results of these variables at a lead time of 72 h, and the tracked cyclone eyes are indicated using the tail of arrows. c ) The procedural tracking results of Typhoon Kong-rey (2018–25). The results of Pangu-Weather were compared to that of ECMWF-HRES and the ground-truth (by IBTrACS 24 , 25 ). We show six time points with the first one being 12:00 UTC, September 29th, 2018, and the time gap between neighboring sub-figures being 12 h. The historical (observed) path of cyclone eyes is shown in dashed. Mind the significant difference between the tracking results of Pangu-Weather and ECMWF-HRES (Pangu-Weather is more accurate) at the middle four sub-figures. The sub-figures with maps were plotted using the Matplotlib Basemap toolkit.

Supplementary information

Peer review file, source data fig. 2, source data fig. 4, source data fig. 5, source data extended data fig. 1, source data extended data fig. 2, source data extended data fig. 3, source data extended data fig. 4, source data extended data fig. 5, source data extended data fig. 7, source data extended data fig. 8, rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Bi, K., Xie, L., Zhang, H. et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature 619 , 533–538 (2023). https://doi.org/10.1038/s41586-023-06185-3

Download citation

Received : 05 January 2023

Accepted : 09 May 2023

Published : 05 July 2023

Issue Date : 20 July 2023

DOI : https://doi.org/10.1038/s41586-023-06185-3

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Cross-layer transmission realized by light-emitting memristor for constructing ultra-deep neural network with transfer learning ability.

• Zhenjia Chen
• Zhenyuan Lin
• Huipeng Chen

Nature Communications (2024)

• Xiaohui Zhong

Weather and climate predicted accurately — without using a supercomputer

• Oliver Watt-Meyer

Nature (2024)

Learning to downscale satellite gravimetry data through artificial intelligence

• Alexander Sun

Nature Water (2024)

WindSeer: real-time volumetric wind prediction over complex terrain aboard a small uncrewed aerial vehicle

• Florian Achermann
• Thomas Stastny
• Nicholas Lawrance

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Help | Advanced Search

Computer Science > Machine Learning

Title: karina: an efficient deep learning model for global weather forecast.

Abstract: Deep learning-based, data-driven models are gaining prevalence in climate research, particularly for global weather prediction. However, training the global weather data at high resolution requires massive computational resources. Therefore, we present a new model named KARINA to overcome the substantial computational demands typical of this field. This model achieves forecasting accuracy comparable to higher-resolution counterparts with significantly less computational resources, requiring only 4 NVIDIA A100 GPUs and less than 12 hours of training. KARINA combines ConvNext, SENet, and Geocyclic Padding to enhance weather forecasting at a 2.5° resolution, which could filter out high-frequency noise. Geocyclic Padding preserves pixels at the lateral boundary of the input image, thereby maintaining atmospheric flow continuity in the spherical Earth. SENet dynamically improves feature response, advancing atmospheric process modeling, particularly in the vertical column process as numerous channels. In this vein, KARINA sets new benchmarks in weather forecasting accuracy, surpassing existing models like the ECMWF S2S reforecasts at a lead time of up to 7 days. Remarkably, KARINA achieved competitive performance even when compared to the recently developed models (Pangu-Weather, GraphCast, ClimaX, and FourCastNet) trained with high-resolution data having 100 times larger pixels. Conclusively, KARINA significantly advances global weather forecasting by efficiently modeling Earth's atmosphere with improved accuracy and resource efficiency.

Submission history

Access paper:.

• HTML (experimental)
• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

• BACKGROUND.
• Primary WRF components.
• Real-time NWP.
• Tailored WRF systems and model applications.
• Teaching and scientific training.
• Community support.
• FUTURE DIRECTIONS.
• CLOSING REMARKS.

Aitken , M. L. , B. Kosović , J. D. Mirocha , and J. K. Lundquist , 2014 : Large eddy simulation of wind turbine wake dynamics in the stable boundary layer using the Weather Research and Forecasting Model . J. Renewable Sustainable Energy , 6 , 033137, doi: 10.1063/1.4885111 .

• Search Google Scholar
• Export Citation

Akter , N ., 2015 : Mesoscale convection and bimodal cyclogenesis over the Bay of Bengal . Mon. Wea. Rev. , 143 , 3495 – 3517 , doi: 10.1175/MWR-D-14-00260.1 .

Alapaty , K. , J. A. Herwehe , T. L. Otte , C. G. Nolte , O. R. Bullock , M. S. Mallard , J. S. Kain , and J. Dudhia , 2012 : Introducing subgrid-scale cloud feedbacks to radiation for regional meteorological and climate modeling . Geophys. Res. Lett. , 39 , L24809, doi: 10.1029/2012GL054031 .

Anderson , J. , T. Hoar , K. Raeder , H. Liu , and N. Collins , 2009 : The Data Assimilation Research Testbed: A community facility . Bull. Amer. Meteor. Soc. , 90 , 1283 – 1296 , doi: 10.1175/2009BAMS2618.1 .

Baklanov , A. , and Coauthors , 2014 : Online coupled regional meteorology chemistry models in Europe: Current status and prospects . Atmos. Chem. Phys. , 14 , 317 – 398 , doi: 10.5194/acp-14-317-2014 .

Barker , D. M. , W. Huang , Y.-R. Guo , A. J. Bourgeois , and Q.-N. Xiao , 2004 : A three-dimensional (3DVAR) data assimilation system for use with MM5: Implementation and initial results . Mon. Wea. Rev. , 132 , 897 – 914 , doi: 10.1175/1520-0493(2004)132<0897:ATVDAS>2.0.CO;2 .

Barker , D. M. , and Coauthors , 2012 : The Weather Research and Forecasting Model’s Community Variational/Ensemble Data Assimilation System . Bull. Amer. Meteor. Soc. , 93 , 831 – 843 , doi: 10.1175/BAMS-D-11-00167.1 .

Barlage , M. , S. Miao , and F. Chen , 2016 : Impact of physics parameterizations on high-resolution weather prediction over two Chinese megacities . J. Geophys. Res. Atmos. , 121 , 4487 – 4498 , doi: 10.1002/2015JD024450 .

Barth , M. C. , and Coauthors , 2015 : The Deep Convective Clouds and Chemistry (DC3) field campaign . Bull. Amer. Meteor. Soc. , 96 , 1281 – 1309 , doi: 10.1175/BAMS-D-13-00290.1 .

Benjamin , S. G. , and Coauthors , 2016 : A North American hourly assimilation and model forecast cycle: The Rapid Refresh . Mon. Wea. Rev. , 144 , 1669 – 1694 , doi: 10.1175/MWR-D-15-0242.1 .

Bleck , R. , 2002 : An oceanic general circulation model framed in hybrid isopycnic-Cartesian coordinates . Ocean Modell. , 4 , 55 – 88 , doi: 10.1016/S1463-5003(01)00012-9 .

Booij , N. , R. C. Ris , and L. H. Holthuijsen , 1999 : A third-generation wave model for coastal regions. Part I: Model description and validation . J. Geophys. Res. , 104 , 7649 – 7666 , doi: 10.1029/98JC02622 .

Brewer , M. C. , C. F. Mass , and B. E. Potter , 2013 : The West Coast thermal trough: Mesoscale evolution and sensitivity to terrain and surface fluxes . Mon. Wea. Rev. , 141 , 2869 – 2896 , doi: 10.1175/MWR-D-12-00305.1 .

Bromwich , D. H. , K. M. Hines , and L.-S. Bai , 2009 : Development and testing of Polar Weather Research and Forecasting model: 2. Arctic Ocean . J. Geophys. Res. , 114 , D08122 , doi: 10.1029/2008JD010300 .

Bromwich , D. H. , F. O. Otieno , K. M. Hines , K. W. Manning , and E. Shilo , 2013 : Comprehensive evaluation of polar weather research and forecasting model performance in the Antarctic . J. Geophys. Res. Atmos. , 118 , 274 – 292 , doi: 10.1029/2012JD018139 .

Bromwich , D. H. , A. B. Wilson , L.-S. Bai , G. W. K. Moore , and P. Bauer , 2016 : A comparison of the regional Arctic System Reanalysis and the global ERA-Interim Reanalysis for the Arctic . Quart. J. Roy. Meteor. Soc. , 142 , 644 – 658 , doi: 10.1002/qj.2527 .

Bruyère , C. L. , J. M. Done , G. J. Holland , and S. Fredrick , 2014 : Bias corrections of global models for regional climate simulations of high-impact weather . Climate Dyn. , 43 , 1847 – 1856 , doi: 10.1007/s00382-013-2011-6 .

Chen , F. , and Coauthors , 2011 : The integrated WRF/urban modelling system: Development, evaluation, and applications to urban environmental problems . Int. J. Climatol. , 31 , 273 – 288 , doi: 10.1002/joc.2158 .

Chen , H. , and D.-L. Zhang , 2013 : On the rapid intensification of Hurricane Wilma (2005). Part II: Convective bursts and the upper-level warm core . J. Atmos. Sci. , 70 , 146 – 162 , doi: 10.1175/JAS-D-12-062.1 .

Chen , H. , D.-L. Zhang , J. Carton , and R. Atlas , 2011 : On the rapid intensification of Hurricane Wilma (2005). Part I: Model prediction and structural changes . Wea. Forecasting , 26 , 885 – 901 , doi: 10.1175/WAF-D-11-00001.1 .

Chen , S. S. , and M. Curcic , 2016 : Ocean surface waves in Hurricane Ike (2008) and Superstorm Sandy (2012): Coupled model predictions and observations . Ocean Modell. , 103 , 161 – 176 , doi: 10.1016/j.ocemod.2015.08.005 .

Chen , S. S. , W. Zhao , M. A. Donelan , and H. L. Tolman , 2013 : Directional wind–wave coupling in fully coupled atmosphere–wave–ocean models: Results from CBLAST-Hurricane . J. Atmos. Sci. , 70 , 3198 – 3215 , doi: 10.1175/JAS-D-12-0157.1 .

Ching , J. , R. Rotunno , M. A. LeMone , A. Martilli , B. Kosović , P. A. Jimenez , and J. Dudhia , 2014 : Convectively induced secondary circulations in fine-grid mesoscale numerical weather prediction models . Mon. Wea. Rev. , 142 , 3284 – 3302 , doi: 10.1175/MWR-D-13-00318.1 .

Clark , A. J. , and Coauthors , 2012 : An overview of the 2010 Hazardous Weather Testbed Experimental Forecast Program Spring Experiment . Bull. Amer. Meteor. Soc. , 93 , 55 – 74 , doi: 10.1175/BAMS-D-11-00040.1 .

Coen , J. L. , M. Cameron , J. Michalakes , E. G. Patton , P. J. Riggan , and K. M. Yedinak , 2013 : WRF-Fire: Coupled weather–wildland fire modeling with the Weather Research and Forecasting model . J. Appl. Meteor. Climatol. , 52 , 16 – 38 , doi: 10.1175/JAMC-D-12-023.1 .

David , C. H. , D. J. Gochis , D. R. Maidment , W. Yu , D. N. Yates , and Z.-L. Yang , 2009 : Using NHDPlus as the land base for the Noah-distributed model . Trans. GIS , 13 , 363 – 377 , doi: 10.1111/j.1467-9671.2009.01169.x .

Davis , C. A. , S. C. Jones , and M. Riemer , 2008a : Hurricane vortex dynamics during Atlantic extratropical transition . J. Atmos. Sci. , 65 , 714 – 736 , doi: 10.1175/2007JAS2488.1 .

Davis , C. A. , and Coauthors , 2008b : Prediction of landfalling hurricanes with the Advanced Hurricane WRF Model . Mon. Wea. Rev. , 136 , 1990 – 2005 , doi: 10.1175/2007MWR2085.1 .

Davis , C. A. , W. Wang , J. Dudhia , and R. Torn , 2010 : Does increased horizontal resolution improve hurricane wind forecasts? Wea. Forecasting , 25 , 1826 – 1841 , doi: 10.1175/2010WAF2222423.1 .

Done , J. M. , G. J. Holland , C. L. Bruyère , L. R. Leung , and A . Suzuki-Parker , 2015 : Modeling high-impact weather and climate: Lessons from a tropical cyclone perspective . Climatic Change , 129 , 381 – 395 , doi: 10.1007/s10584-013-0954-6 .

Donelan , M. A. , M. Curcic , S. S. Chen , and A. K. Magnusson , 2012 : Modeling waves and wind stress . J. Geophys. Res. , 117 , C00J23 , doi: 10.1029/2011JC007787 .

DTC/NCEP , 2014 : NMM version 3 modeling system user’s guide. Developmental Testbed Center, chap. 1–7. [Available online at www.dtcenter.org/wrf-nmm/users/docs/user_guide/V3/users_guide_nmm_chap1-7.pdf .]

Dudhia , J ., 2014 : A history of mesoscale model development . Asia-Pac. J. Atmos. Sci. , 50 , 121 – 131 , doi: 10.1007/s13143-014-0031-8 .

DuVivier , A. K. , and J. J. Cassano , 2013 : Evaluation of WRF Model resolution on simulated mesoscale winds and surface fluxes near Greenland . Mon. Wea. Rev. , 141 , 941 – 963 , doi: 10.1175/MWR-D-12-00091.1 .

Fang , J. , and F. Zhang , 2010 : Initial development and genesis of Hurricane Dolly (2008) . J. Atmos. Sci. , 67 , 655 – 672 , doi: 10.1175/2009JAS3115.1 .

Fast , J. D. , W. I. Gustafson , E. G. Chapman , R. C. Easter , J. Rishel , R. A. Zaveri , G. A. Grell , and M. Barth , 2006 : Evolution of ozone, particulates, and aerosol direct forcing in an urban area using a new fully-coupled meteorology, chemistry, and aerosol model . J. Geophys. Res. , 111 , D21305 , doi: 10.1029/2005JD006721 .

Fierro , A. O. , R. F. Rogers , F. D. Marks , and D. S. Nolan , 2009 : The impact of horizontal grid spacing on the microphysical and kinematic structures of strong tropical cyclones simulated with the WRF-ARW model . Mon. Wea. Rev. , 137 , 3717 – 3743 , doi: 10.1175/2009MWR2946.1 .

Ganetis , S. A. , and B. A. Colle , 2015 : The thermodynamic and microphysical evolution of an intense snowband during the Northeast U.S. blizzard of 8–9 February 2013 . Mon. Wea. Rev. , 143 , 4104 – 4125 , doi: 10.1175/MWR-D-14-00407.1 .

Givati , A. , D. J. Gochis , T. Rummler , and H. Kunstmann , 2016 : Comparing one-way and two-way coupled hydrometeorological forecasting systems for flood forecasting in the Mediterranean region . Hydrology , 3 , 19 , doi: 10.3390/hydrology3020019 .

Gochis , D. J. , W. Yu , and D. N. Yates , 2015 : The WRF-Hydro Model technical description and user’s guide, version 3.0. NCAR Tech. Doc., 123 pp. [Available online at https://ral.ucar.edu/sites/default/files/public/images/project/WRF_Hydro_User_Guide_v3.0.pdf .]

Gopalakrishnan , S. G. , S. Goldenberg , T. Quirino , X. Zhang , F. Marks Jr. , K.-S. Yeh , R. Atlas , and V. Tallapragada , 2012 : Toward improving high-resolution numerical hurricane forecasting: Influence of model horizontal grid resolution, initialization, and physics . Wea. Forecasting , 27 , 647 – 666 , doi: 10.1175/WAF-D-11-00055.1 .

Grell , G. A. , and A. Baklanov , 2011 : Integrated modeling for forecasting weather and air quality: A call for fully coupled approaches . Atmos. Environ. , 45 , 6845 – 6851 , doi: 10.1016/j.atmosenv.2011.01.017 .

Grell , G. A. , J. Dudhia , and D. R. Stauffer , 1994 : A description of the fifth-generation Penn State/NCAR Mesoscale Model (MM5). NCAR Tech. Note TN-398+STR, 121 pp., doi: 10.5065/D60Z716B .

Grell , G. A. , R. Knoche , S. E. Peckham , and S. McKeen , 2004 : Online versus offline air quality modeling on cloud-resolving scales . Geophys. Res. Lett. , 31 , L16117 , doi: 10.1029/2004GL020175 .

Grell , G. A. , R. Knoche , R. Schmitz , S. A. McKeen , G. Frost , W. C. Skamarock , and B. Eder , 2005 : Fully-coupled “online” chemistry within the WRF model . Atmos. Environ. , 39 , 6957 – 6976 , doi: 10.1016/j.atmosenv.2005.04.027 .

Haupt , S. E. , and W. P. Mahoney , 2015 : Taming wind power with better forecasts . IEEE Spectrum , 52 ( 11 ), 47 – 52 , doi: 10.1109/MSPEC.2015.7335902 .

Hines , K. M. , and D. H. Bromwich , 2008 : Development and testing of Polar Weather Research and Forecasting (WRF) Model. Part I: Greenland ice sheet meteorology . Mon. Wea. Rev. , 136 , 1971 – 1989 , doi: 10.1175/2007MWR2112.1 .

Hines , K. M. , D. H. Bromwich , L.-S. Bai , M. Barlage , and A. G. Slater , 2011 : Development and testing of Polar WRF. Part III: Arctic land . J. Climate , 24 , 26 – 48 , doi: 10.1175/2010JCLI3460.1 .

Hong , S.-Y. , and J. Dudhia , 2012 : Next-generation numerical weather prediction: Bridging parameterization, explicit clouds, and large eddies . Bull. Amer. Meteor. Soc. , 93 , ES6 – ES9 , doi: 10.1175/2011BAMS3224.1 .

Hsiao , L. F. , D.-S. Chen , Y.-H. Kuo , Y.-R. Guo , T. C. Yeh , J.-S. Hong , C.-T. Fong , and C.-S. Lee , 2012 : Application of WRF 3DVAR to operational typhoon prediction in Taiwan: Impact of outer loop and partial cycling approaches . Wea. Forecasting , 27 , 1249 – 1263 , doi: 10.1175/WAF-D-11-00131.1 .

Huang , M. , B. Huang , and A. H.-L. Huang , 2014 : Application of Intel Many Integrated Core (MIC) architecture to the Yonsei University planetary boundary layer scheme in the Weather Research and Forecasting model. High-Performance Computing in Remote Sensing IV , B. Huang, S. López, and Z. Wu, Eds., International Society for Optical Engineering (SPIE Proceedings, Vol. 9247), 92470J, doi: 10.1117/12.2069430 .

Huang , X.-Y. , and Coauthors , 2009 : Four-dimensional variational data assimilation for WRF: Formulation and preliminary results . Mon. Wea. Rev. , 137 , 299 – 314 , doi: 10.1175/2008MWR2577.1 .

Hurrell , J. W. , and Coauthors , 2013 : The Community Earth System Model: A framework for collaborative research . Bull. Amer. Meteor. Soc. , 94 , 1339 – 1360 , doi: 10.1175/BAMS-D-12-00121.1 .

Hutchinson , T. A. , 2015 : Global WRF-based forecast system: Implementation and applications. 43rd Conf. on Broadcast Meteorology , Raleigh, NC, Amer. Meteor. Soc., 6.5. [Available online at https://ams.confex.com/ams/43BC3WxWarn/webprogram/Paper271512.html .]

Janjić , Z. I. , 2003 : A nonhydrostatic model based on a new approach . Meteor. Atmos. Phys. , 82 , 271 – 285 , doi: 10.1007/s00703-001-0587-6 .

Janjić , Z. I. , J. P. Gerrity Jr. , and S. Nikovic , 2001 : An alternative approach to nonhydrostatic modeling . Mon. Wea. Rev. , 129 , 1164 – 1178 , doi: 10.1175/1520-0493(2001)129<1164:AAATNM>2.0.CO;2 .

Jimenez , P. , and Coauthors , 2016 : WRF-Solar: An augmented NWP model for solar power prediction. Model description and clear sky assessment . Bull. Amer. Meteor. Soc. , 97 , 1249 – 1264 , doi: 10.1175/BAMS-D-14-00279.1 .

Jin , C.-S. , C.-H. Ho , J.-H. Kim , D.-K. Lee , D.-H. Cha , and S.-W. Yeh , 2013 : Critical role of northern off-equatorial sea surface temperature forcing associated with central Pacific El Niño in more frequent tropical cyclone movements toward East Asia . J. Climate , 26 , 2534 – 2545 , doi: 10.1175/JCLI-D-12-00287.1 .

Jung , B.-J. , H. Y. Kim , T. Auligné , X. Zhang , X. Zhang , and X.-Y. Huang , 2013 : Adjoint-derived observation impact using WRF in the western North Pacific . Mon. Wea. Rev. , 141 , 4080 – 4097 , doi: 10.1175/MWR-D-12-00197.1 .

Karamchandani , P. , Y. Zhang , S.-Y. Chen , and R. Balmori-Bronson , 2012 : Development of an extended chemical mechanism for global-through-urban applications . Atmos. Pollut. Res. , 3 , 1 – 24 , doi: 10.5094/APR.2011.047 .

Khain , A. , B. Lynn , and J. Dudhia , 2010 : Aerosol effects on intensity of landfalling hurricanes as seen from simulations with the WRF Model with spectral bin microphysics . J. Atmos. Sci. , 67 , 365 – 384 , doi: 10.1175/2009JAS3210.1 .

Kim , H.-Y. , C. Lozano , V. Tallapragada , D. Iredell , D. Sheinin , H. L. Tolman , V. M. Gerald , and J. Sims , 2014 : Performance of ocean simulations in the coupled HWRF–HYCOM Model . J. Atmos. Oceanic Technol. , 31 , 545 – 559 , doi: 10.1175/JTECH-D-13-00013.1 .

Kirkil , G. , J. Mirocha , E. Bou-Zeid , F. K. Chow , and B. Kosović , 2012 : Implementation and evaluation of dynamic subfilter-scale stress models for large-eddy simulation using WRF . Mon. Wea. Rev. , 140 , 266 – 284 , doi: 10.1175/MWR-D-11-00037.1 .

Kleist , D. T. , D. F. Parrish , J. C. Derber , R. Treadon , W.-S. Wu , and S. Lord , 2009 : Introduction of the GSI into the NCEP Global Data Assimilation System . Wea. Forecasting , 24 , 1691 – 1705 , doi: 10.1175/2009WAF2222201.1 .

Klemp , J. B. , W. C. Skamarock , and J. Dudhia , 2007 : Conservative split-explicit time integration methods for the compressible nonhydrostatic equations . Mon. Wea. Rev. , 135 , 2897 – 2913 , doi: 10.1175/MWR3440.1 .

Kochanski , A. K. , M. A. Jenkins , K. Yedinak , J. Mandel , J. Beezley , and B. Lamb , 2015 : Toward an integrated system for fire, smoke and air quality simulations . Int. J. Wildland Fire , 25 , 534 – 546 , doi: 10.1071/WF14074 .

Leung , L. R. , and Y. Qian , 2009 : Atmospheric rivers induced heavy precipitation and flooding in the western U.S. simulated by the WRF regional climate model . Geophys. Res. Lett. , 36 , L03820 , doi: 10.1029/2008GL036445 .

Liu , B. , H. Liu , L. Xie , C. Guan , and D. Zhao , 2011 : A coupled atmosphere–wave–ocean modeling system: Simulation of the intensity of an idealized tropical cyclone . Mon. Wea. Rev. , 139 , 132 – 152 , doi: 10.1175/2010MWR3396.1 .

Liu , N. , T. Ling , H. Wang , Y. Zhang , Z. Gao , and Y. Wang , 2015 : Numerical simulation of Typhoon Muifa (2011) using a Coupled Ocean-Atmosphere-Wave-Sediment Transport (COAWST) modeling system . J. Ocean Univ. China , 14 , 199 – 209 , doi: 10.1007/s11802-015-2415-5 .

Liu , Y. , and Coauthors , 2011 : Simultaneous nested modeling from the synoptic scale to the LES scale for wind energy applications . J. Wind Eng. Ind. Aerodyn. , 99 , 308 – 319 , doi: 10.1016/j.jweia.2011.01.013 .

Lu , Y. , and Y. Deng , 2015 : Initial transient response of an intensifying baroclinic wave to increases in cloud droplet number concentration . J. Climate , 28 , 9669 – 9677 , doi: 10.1175/JCLI-D-15-0251.1 .

Mahoney , W. P. , and Coauthors , 2012 : A wind power forecasting system to optimize grid integration . IEEE Trans. Sustainable Energy , 3 , 670 – 682 , doi: 10.1109/TSTE.2012.2201758 .

Mass , C. F. , M. D. Warner , and R. Steed , 2014 : Strong westerly wind events in the Strait of Juan de Fuca . Wea. Forecasting , 29 , 445 – 465 , doi: 10.1175/WAF-D-13-00026.1 .

Mellor , G. L. , 2004 : Users guide for a three-dimensional, primitive equation, numerical ocean model. Princeton University, 56 pp.

Meng , Z. , F. Zhang , P. Markowski , D. Wu , and K. Zhao , 2012 : A modeling study on the development of a bowing structure and associated rear inflow within a squall line over south China . J. Atmos. Sci. , 69 , 1182 – 1207 , doi: 10.1175/JAS-D-11-0121.1 .

Michalakes , J. G. , J. Dudhia , D. O. Gill , J. B. Klemp , and W. C. Skamarock , 1999 : Design of a next-generation regional weather research and forecast model. Towards Teracomputing: Proceedings of the Eighth ECMWF Workshop on the Use of Parallel Processors in Meteorology , W. Zweiflhofer and N. Kreitz, Eds., World Scientific, 117–124.

Mielikainen , J. , B. Huang , and A. H.-L. Huang , 2014 : Intel Xeon Phi accelerated Weather Research and Forecasting (WRF) Goddard microphysics scheme . Geosci. Model. Dev. Discuss. , 7 , 8941 – 8973 , doi: 10.5194/gmdd-7-8941-2014 .

Miller , W. , H. Chen , and D.-L. Zhang , 2015 : On the rapid intensification of Hurricane Wilma (2005). Part III: Effects of latent heat of fusion . J. Atmos. Sci. , 72 , 3829 – 3849 , doi: 10.1175/JAS-D-14-0386.1 .

Mirocha , J. D. , J. K. Lundquist , and B. Kosović , 2010 : Implementation of a nonlinear subfilter turbulence stress model for large-eddy simulation in the Advanced Research WRF Model . Mon. Wea. Rev. , 138 , 4212 – 4228 , doi: 10.1175/2010MWR3286.1 .

Mirocha , J. D. , B. Kosović , M. L. Aitken , and J. K. Lundquist , 2014 : Implementation of a generalized actuator disk wind turbine model into the weather research and forecasting model for large-eddy simulation applications . J. Renewable Sustainable Energy , 6 , 013104, doi: 10.1063/1.4861061 .

Moeng , C.-H. , J. Dudhia , J. B. Klemp , and P. Sullivan , 2007 : Examining two-way grid nesting for large eddy simulation of the PBL using the WRF model . Mon. Wea. Rev. , 135 , 2295 – 2311 , doi: 10.1175/MWR3406.1 .

Molinari , J. , and M. Dudek , 1992 : Parameterization of convective precipitation in mesoscale numerical models: A critical review . Mon. Wea. Rev. , 120 , 326 – 344 , doi: 10.1175/1520-0493(1992)120<0326:POCPIM>2.0.CO;2 .

Moon , Y. , and D. S. Nolan , 2015 : Spiral rainbands in a numerical simulation of Hurricane Bill (2009). Part I: Structures and comparisons to observations . J. Atmos. Sci. , 72 , 164 – 190 , doi: 10.1175/JAS-D-14-0058.1 .

Morrison , H. , S. A. Tessendorf , K. Ikeda , and G. Thompson , 2012 : Sensitivity of a simulated midlatitude squall line to parameterization of raindrop breakup . Mon. Wea. Rev. , 140 , 2437 – 2460 , doi: 10.1175/MWR-D-11-00283.1 .

Muñoz-Esparza , D. , B. Kosović , J. D. Mirocha , and J. van Beeck , 2014 : Bridging the transition from mesoscale to microscale turbulence in numerical weather prediction models . Bound.-Layer Meteor. , 153 , 409 – 440 , doi: 10.1007/s10546-014-9956-9 .

Muñoz-Esparza , D. , B. Kosović , J. D. Mirocha , and J. van Beeck , 2015 : A stochastic perturbation method to generate inflow turbulence in large-eddy simulation models: Application to neutrally stratified atmospheric boundary layers . Phys. Fluids , 27 , 035102, doi: 10.1063/1.4913572 .

NCAR , 2016 : Information for WRF code contributors. [Available online at http://www2.mmm.ucar.edu/wrf/users/contrib_info.php .]

NCAR , 2017 : User's guide for the Advanced Research WRF (ARW) modeling system, Version 3. National Center for Atmospheric Research, 443 pp. [Available online at http://www2.mmm.ucar.edu/wrf/users/docs/user_guide_V3/contents.html .]

NOAA , 2016 : WRF-CHEM version 3.8.1 user’s guide. 73 pp. [Available online at https://ruc.noaa.gov/wrf/wrf-chem/Users_guide.pdf .]

Parish , T. R. , D. A. Rahn , and D. Leon , 2015 : Aircraft observations and numerical simulations of the developing stage of a southerly surge near Southern California . Mon. Wea. Rev. , 143 , 4883 – 4903 , doi: 10.1175/MWR-D-15-0140.1 .

Park , S.-H. , J. B. Klemp , and W. C. Skamarock , 2014 : A comparison of mesh refinement in the global MPAS-A and WRF models using an idealized normal-mode baroclinic wave simulation . Mon. Wea. Rev. , 142 , 3614 – 3634 , doi: 10.1175/MWR-D-14-00004.1 .

Peace , M. , T. Mattner , G. Mills , J. Kepert , and L. McCaw , 2015 : Fire-modified meteorology in a coupled fire–atmosphere model . J. Appl. Meteor. Climatol. , 54 , 704 – 720 , doi: 10.1175/JAMC-D-14-0063.1 .

Peckham , S. E. , T. G. Smirnova , S. G. Benjamin , J. M. Brown , and J. S. Kenyon , 2016 : Implementation of a digital filter initialization in the WRF Model and its application in the Rapid Refresh . Mon. Wea. Rev. , 144 , 99 – 106 , doi: 10.1175/MWR-D-15-0219.1 .

Powers , J. G. , 2007 : Numerical prediction of an Antarctic severe wind event with the Weather Research and Forecasting (WRF) Model . Mon. Wea. Rev. , 135 , 3134 – 3157 , doi: 10.1175/MWR3459.1 .

Powers , J. G. , K. W. Manning , D. H. Bromwich , J. J. Cassano , and A. M. Cayette , 2012 : A decade of Antarctic science support through AMPS . Bull. Amer. Meteor. Soc. , 93 , 1699 – 1712 , doi: 10.1175/BAMS-D-11-00186.1 .

Richardson , M. I. , A. D. Toigo , and C. E. Newman , 2007 : PlanetWRF: A general purpose, local to global numerical model for planetary atmosphere and climate dynamics . J. Geophys. Res. , 112 , E09001 , doi: 10.1029/2006JE002825 .

Romine , G. S. , C. S. Schwartz , J. Berner , K. R. Fossell , C. M. Snyder , J. L. Anderson , and M. L. Weisman , 2014 : Representing forecast error in a convection-permitting ensemble system . Mon. Wea. Rev. , 142 , 4519 – 4541 , doi: 10.1175/MWR-D-14-00100.1 .

Romine , G. S. , C. S. Schwartz , R. D. Torn , and M. L. Weisman , 2016 : Impact of assimilating dropsonde observations from MPEX on ensemble forecasts of severe weather events . Mon. Wea. Rev. , 144 , 3799 – 3823 , doi: 10.1175/MWR-D-15-0407.1 .

Rostom , R. , and Y. L. Lin , 2015 : Control parameters for track continuity of cyclones passing over the south-central Appalachian Mountains . Wea. Forecasting , 30 , 1429 – 1449 , doi: 10.1175/WAF-D-14-00080.1 .

Satish , N. , C. Kim , J. Chugani , A. Nguyen , V. Lee , D. Kim , and P. Dubey , 2010 : Fast sort on CPUs, GPUs and Intel MIC architectures. Intel Tech. Rep., 11 pp.

Schultz , D. M. , and J. M. Sienkiewicz , 2013 : Using frontogenesis to identify sting jets in extratropical cyclones . Wea. Forecasting , 28 , 603 – 613 , doi: 10.1175/WAF-D-12-00126.1 .

Schwartz , C. S. , G. S. Romine , M. L. Weisman , R. A. Sobash , K. R. Fossell , K. W. Manning , and S. B. Trier , 2015a : A real-time convection-allowing ensemble prediction system initialized by mesoscale ensemble Kalman filter analyses . Wea. Forecasting , 30 , 1158 – 1181 , doi: 10.1175/WAF-D-15-0013.1 .

Schwartz , C. S. , G. S. Romine , R. A. Sobash , K. R. Fossell , and M. L. Weisman , 2015b : NCAR’s experimental real-time convection-allowing ensemble prediction system . Wea. Forecasting , 30 , 1645 – 1654 , doi: 10.1175/WAF-D-15-0103.1 .

Schwartz , C. S. , Z. Liu , and X.-Y. Huang , 2015c : Sensitivity of limited-area hybrid variational-ensemble analyses and forecasts to ensemble perturbation resolution . Mon. Wea. Rev. , 143 , 3454 – 3477 , doi: 10.1175/MWR-D-14-00259.1 .

Senatore , A. , G. Mendicino , D. J. Gochis , W. Yu , D. N. Yates , and H. Kunstmann , 2015 : Fully coupled atmosphere-hydrology simulations for the central Mediterranean: Impact of enhanced hydrological parameterization for short and long time scales . J. Adv. Model. Earth Syst. , 7 , 1693 – 1715 , doi: 10.1002/2015MS000510 .

Shchepetkin , A. F. , and J. C. McWilliams , 2005 : The Regional Ocean Modeling System: A split-explicit, free-surface, topography-following coordinates ocean model . Ocean Modell. , 9 , 347 – 404 , doi: 10.1016/j.ocemod.2004.08.002 .

Shi , J. J. , and Coauthors , 2010 : WRF simulations of the 20–22 January 2007 snow events over eastern Canada: Comparison with in situ and satellite observations . J. Appl. Meteor. Climatol. , 49 , 2246 – 2266 , doi: 10.1175/2010JAMC2282.1 .

Shin , H. H. , and S.-Y. Hong , 2013 : Analysis of resolved and parameterized vertical transports in convective boundary layers at gray-zone resolutions . J. Atmos. Sci. , 70 , 3248 – 3261 , doi: 10.1175/JAS-D-12-0290.1 .

Shin , H. H. , and S.-Y. Hong , 2015 : Representation of the subgrid-scale turbulent transport in convective boundary layers at gray-zone resolutions . Mon. Wea. Rev. , 143 , 250 – 271 , doi: 10.1175/MWR-D-14-00116.1 .

Simpson , C. C. , J. J. Sharples , J. P. Evans , and M. F. McCabe , 2013 : Large eddy simulation of atypical wildland fire spread on leeward slopes . Int. J. Wildland Fire , 22 , 599 – 614 , doi: 10.1071/WF12072 .

Simpson , C. C. , J. J. Sharples and J. P. Evans , 2014 : Resolving vorticity-driven lateral fire spread using the WRF-Fire coupled atmosphere–fire numerical model . Nat. Hazards Earth Syst. Sci. , 14 , 2359 – 2371 , doi: 10.5194/nhess-14-2359-2014 .

Skamarock , W. C. , and Coauthors , 2008 : A description of the Advanced Research WRF version 3. NCAR Tech. Note NCAR/TN-475+STR, 113 pp., doi: 10.5065/D68S4MVH .

Skamarock , W. C. , J. B. , Klemp , M. G. Duda , L. Fowler , S.-H. Park , and T. D. Ringler , 2012 : A multiscale nonhydrostatic atmospheric model using centroidal Voronoi tessellations and C-grid staggering . Mon. Wea. Rev. , 140 , 3090 – 3105 , doi: 10.1175/MWR-D-11-00215.1 .

Smith , J. , 2015 : Saving farmers money in tropical West Africa. UCAR. [Available online at https://www2.ucar.edu/atmosnews/in-brief/17896/saving-farmers-money-tropical-west-africa .]

Talbot , C. , E. Bou-Zeid , and J. Smith , 2012 : Nested mesoscale large-eddy simulations with WRF: Performance in real test cases . J. Hydrometeor. , 13 , 1421 – 1441 , doi: 10.1175/JHM-D-11-048.1 .

Tallapragada , V. , and Coauthors , 2014 : Hurricane Weather Research and Forecasting (HWRF) Model: 2014 scientific documentation. HWRF v3.6a, NCAR Developmental Testbed Center Rep., 105 pp.

Thompson , G. , and T. Eidhammer , 2014 : A study of aerosol impacts on clouds and precipitation development in a large winter cyclone . J. Atmos. Sci. , 71 , 3636 – 3658 , doi: 10.1175/JAS-D-13-0305.1 .

Thompson , G. , M. Tewari , K. Ikeda , S. Tessendorf , C. Weeks , J. Otkin , and F. Kong , 2016 : Explicitly-coupled cloud physics and radiation parameterizations and subsequent evaluation in WRF high-resolution convective forecasts . Atmos. Res. , 168 , 92 – 104 , doi: 10.1016/j.atmosres.2015.09.005 .

Tolman , H. L. , 2009 : User manual and system documentation of WAVEWATCH III, version 3.14. NOAA/NCEP/EMC/MMAB Tech. Note 276, 220 pp. [Available from NOAA National Centers for Environmental Prediction, Environmental Modeling Center, Marine Modeling and Analysis Branch, 5830 University Research Ct., College Park, MD 20740.]

Trier , S. B. , C. A. Davis , D. A. Ahijevych , M. L. Weisman , and G. H. Bryan , 2006 : Mechanisms supporting long-lived episodes of propagating nocturnal convection within a 7-day WRF Model simulation . J. Atmos. Sci. , 63 , 2437 – 2461 , doi: 10.1175/JAS3768.1 .

Volkamer , R. , and Coauthors , 2015 : Aircraft measurements of BrO, IO, glyoxal, NO 2 , H 2 O, O 2 –O 2 , and aerosol extinction profiles in the tropics: Comparison with aircraft-/ship-based in situ and lidar measurements . Atmos. Meas. Tech. , 8 , 2121 – 2148 , doi: 10.5194/amt-8-2121-2015 .

Wagner , N. L. , and Coauthors , 2015 : In situ vertical profiles of aerosol extinction, mass, and composition over the southeast United States during SENEX and SEAC4RS: Observations of a modest aerosol enhancement aloft . Atmos. Chem. Phys. , 15 , 7085 – 7102 , doi: 10.5194/acp-15-7085-2015 .

Wang , X. , D. M. Barker , C. M. Snyder , and T. M. Hamill , 2008 : A hybrid ETKF-3DVAR data assimilation scheme for the WRF Model. Part I: Observing system simulation experiment . Mon. Wea. Rev. , 136 , 5116 – 5131 , doi: 10.1175/2008MWR2444.1 .

Wang , Z. , M. T. Montgomery , and T. J. Dunkerton , 2010 : Genesis of pre–Hurricane Felix (2007). Part II: Warm core formation, precipitation evolution, and predictability . J. Atmos. Sci. , 67 , 1730 – 1744 , doi: 10.1175/2010JAS3435.1 .

Warner , J. C. , C. R. Sherwood , R. P. Signell , C. Harris , and H. G. Arango , 2008 : Development of a three-dimensional, regional, coupled wave, current, and sediment-transport model . Comput. Geosci. , 34 , 1284 – 1306 , doi: 10.1016/j.cageo.2008.02.012 .

Warner , J. C. , B. Armstrong , R. He , and J. Zambon , 2010 : Development of a Coupled Ocean–Atmosphere–Wave–Sediment Transport (COAWST) Modeling System . Ocean Modell. , 35 , 230 – 244 , doi: 10.1016/j.ocemod.2010.07.010 .

Weisman , M. L. , and Coauthors , 2015 : The Mesoscale Predictability Experiment (MPEX) . Bull. Amer. Meteor. Soc. , 96 , 2127 – 2149 , doi: 10.1175/BAMS-D-13-00281.1 .

Wu , W.-S. , R. J. Purser , and D. F. Parrish , 2002 : Three-dimensional variational analysis with spatially inhomogeneous covariances . Mon. Wea. Rev. , 130 , 2905 – 2916 , doi: 10.1175/1520-0493(2002)130<2905:TDVAWS>2.0.CO;2 .

Wyngaard , J. C. , 2004 : Toward numerical modeling in the “Terra Incognita.” J. Atmos. Sci. , 61 , 1816 – 1826 , doi: 10.1175/1520-0469(2004)061<1816:TNMITT>2.0.CO;2 .

Xiao , H. , and Coauthors , 2015 : Modifications to WRF’s dynamical core to improve the treatment of moisture for large-eddy simulations . J. Adv. Model. Earth Syst. , 7 , 1627 – 1642 , doi: 10.1002/2015MS000532 .

Xu , D. , Z. Liu , X.-Y. Huang , J. Min , and H. Wang , 2013 : Impact of assimilating IASI radiance observations on forecasts of two tropical cyclones . Meteor. Atmos. Phys. , 122 , 1 – 18 , doi: 10.1007/s00703-013-0276-2 .

Xu , X. , M. Xue , and Y. Wang , 2015 : The genesis of mesovortices within a real-data simulation of a bow echo system . J. Atmos. Sci. , 72 , 1963 – 1986 , doi: 10.1175/JAS-D-14-0209.1 .

Yu , X. , and T.-Y. Lee , 2010 : Role of convective parameterization in simulations of a convection band at grey-zone resolutions . Tellus , 62A , 617 – 632 , doi: 10.1111/j.1600-0870.2010.00470.x .

Yucel , I. , A. Onen , K. Yilmaz , and D. J. Gochis , 2015 : Calibration and evaluation of a flood forecasting system: Utility of numerical weather prediction model, data assimilation and satellite-based rainfall . J. Hydrol. , 523 , 49 – 66 , doi: 10.1016/j.jhydrol.2015.01.042 .

Zambon , J. B. , R. He , and J. C. Warner , 2014 : Investigation of Hurricane Ivan using the coupled ocean–atmosphere–wave–sediment transport (COAWST) model . Ocean Dyn. , 64 , 1535 – 1554 , doi: 10.1007/s10236-014-0777-7 .

Zhang , Y. , J. A. Smith , A. A. Ntelekos , M. L. Baeck , W. F. Krajewski , and F. Moshary , 2009 : Structure and evolution of precipitation along a cold front in the northeastern United States . J. Hydrometeor. , 10 , 1243 – 1256 , doi: 10.1175/2009JHM1046.1 .

Zhang , Y. , J. Hemperly , N. Meskhidze , and W. C. Skamarock , 2012 : The Global Weather Research and Forecasting Model: Model evaluation, sensitivity study, and future year simulation . Atmos. Climate Sci. , 2 , 231 – 253 , doi: 10.4236/acs.2012.23024 .

Zhu , P ., 2008 : Simulation and parameterization of the turbulent transport in the hurricane boundary layer by large eddies . J. Geophys. Res. , 113 , D17104 , doi: 10.1029/2007JD009643 .

(left) WRF registrations and (right) WRF publications through 2015. In the left panel, cumulative (red) and annual (blue) registrations since initial release. In the right panel, cumulative (red) and annual (blue) WRF-related publications.

Countries that have logged registered WRF users (gold) and that have logged users and have run WRF operationally (orange). In some countries, the WRF operation has been in regional meteorological centers in selected cities. Note also that operational centers may have run multiple NWP models, with WRF not being the exclusive model.

Squall line event of 12–13 Jul 2015 in the central United States with WRF forecast. (left) NOAA Storm Prediction Center (SPC) image of composited hourly radar reflectivity for a squall line traveling from Minnesota to Ohio, 12–13 Jul 2015. Individual positions of the squall line are marked by enhanced reflectivity in red/orange, and the line progresses from northwest to southeast (from Minnesota to Kentucky). Approximate line lifetime and period captured in image is 2300 UTC 12 Jul–1800 UTC 13 Jul (Image: Greg Carbin, SPC). (right) Composited hourly simulated reflectivity from a WRF forecast from the real-time MMM ensemble for 0200 UTC 12 Jul–2100 UTC 13 Jul. Hours 26–45 from forecast initialized at 0000 UTC 12 Jul 2015 are shown (Image: Ryan Sobash, NCAR).

Example of an HRRR forecast (12 h), based on its 3-km CONUS grid, with observed radar reflectivity (dB Z ). (top) Simulated reflectivity from HRRR (WRF) 12-h forecast valid 0000 UTC 21 Apr 2016 (1200 UTC 20 Apr initialization). (bottom) Composited radar reflectivity for 0000 UTC 21 Apr 2016.

Schematic of process components of the WRF-Hydro modeling system and path to production of full hydrologic predictions.

Depiction of WRF-Fire output from a 200-m simulation of a grassland fire in Colorado in 2015, with satellite imagery of the surface underlain. Output represents hour 14 of a forecast initialized at 1200 UTC 18 Sep 2015. Vectors [plotted every fourth grid point (i.e., 800 m)] show model surface wind flow. Colored ring marks fire, with sensible heat flux (W m –2 ) shaded. Scale is in the lower-right corner. Gray shading approximates smoke, based on the WRF-Fire forecast potential temperature tendency.

• View raw image
• Download Powerpoint Slide

Related Content

The impacts of interannual climate variability on the declining trend in terrestrial water storage over the tigris–euphrates river basin, evaluation of snowfall retrieval performance of gpm constellation radiometers relative to spaceborne radars, a 440-year reconstruction of heavy precipitation in california from blue oak tree rings, quantifying the role of internal climate variability and its translation from climate variables to hydropower production at basin scale in india, extreme convective rainfall and flooding from winter season extratropical cyclones in the mid-atlantic region of the united states.

• Previous Article
• Next Article

The Weather Research and Forecasting Model: Overview, System Efforts, and Future Directions

Displayed acceptance dates for articles published prior to 2023 are approximate to within a week. If needed, exact acceptance dates can be obtained by emailing  [email protected] .

• Download PDF

Since its initial release in 2000, the Weather Research and Forecasting (WRF) Model has become one of the world’s most widely used numerical weather prediction models. Designed to serve both research and operational needs, it has grown to offer a spectrum of options and capabilities for a wide range of applications. In addition, it underlies a number of tailored systems that address Earth system modeling beyond weather. While the WRF Model has a centralized support effort, it has become a truly community model, driven by the developments and contributions of an active worldwide user base. The WRF Model sees significant use for operational forecasting, and its research implementations are pushing the boundaries of finescale atmospheric simulation. Future model directions include developments in physics, exploiting emerging compute technologies, and ever-innovative applications. From its contributions to research, forecasting, educational, and commercial efforts worldwide, the WRF Model has made a significant mark on numerical weather prediction and atmospheric science.

© 2017 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy ( www.ametsoc.org/PUBSReuseLicenses ).

As arguably the world’s most widely used numerical weather prediction model, the Weather Research and Forecasting Model offers a spectrum of capabilities for an extensive range of applications.

The Weather Research and Forecasting (WRF) Model ( Skamarock et al. 2008 ) is an atmospheric model designed, as its name indicates, for both research and numerical weather prediction (NWP). While it is officially supported by the National Center for Atmospheric Research (NCAR), WRF has become a true community model by its long-term development through the interests and contributions of a worldwide user base. From these, WRF has grown to provide specialty capabilities for a range of Earth system prediction applications, such as air chemistry, hydrology, wildland fires, hurricanes, and regional climate. The software framework of WRF has facilitated such extensions and supports efficient, massively-parallel computation across a broad range of computing platforms. As detailed below, this paper aims to provide a review of the WRF system and to convey to the meteorological community its significance via its contributions to atmospheric science and weather prediction.

Since its initial public release in 2000, WRF has become arguably the world’s most-used atmospheric model. This is evidenced in metrics of registered users and publications. For example, the cumulative number of WRF registrations is now over 36,000, 1 distributed across 162 countries. Figure 1 (left) shows WRF’s steady growth in cumulative registrations since its initial release, while Fig. 2 maps the countries that have logged registered users (as well as those that have had operational forecasting users). WRF’s widespread acceptance is in part due to its being provided without cost, copyright encumbrance, or restrictions on modification. A measure of the ongoing interest in WRF and the influx of users is the number of annual registrations ( Fig. 1 , left). These averaged over 3,600 per year in the five years from 2011 to 2015. Meanwhile 8,200 users subscribe to the wrf-news e-mail for model information and updates.

Citation: Bulletin of the American Meteorological Society 98, 8; 10.1175/BAMS-D-15-00308.1

• Download Figure
• Download figure as PowerPoint slide

The catalog of WRF-related publications reflects the model’s impact on science. The number of peer-reviewed journal publications involving WRF is currently over 3,500, and the annual average for the 2011–15 period is 510 per year ( Fig. 1 , right). The number of unique institutions on peer-reviewed WRF publications is over 1,340, and the number of unique authors exceeds 11,700. To date, the number of citations to WRF papers is over 26,500, with an average of over 10 citations per publication.

Though WRF is mature, it continues to advance. The system is being vigorously applied in new research directions, real-time settings, and marketplace opportunities. In light of WRF’s continuing growth, its prominent role in research, and its extensive use in forecasting, this article seeks to review this major NWP capability in order to inform and update the meteorological community on the current scope of the system, how it is being applied, and where it is going.

During the 1990s the fifth-generation Pennsylvania State University–NCAR Mesoscale Model (MM5; Grell et al. 1994 ) saw widespread use in the research community. This stemmed largely from its abilities to address increasingly smaller atmospheric scales and to operate on workstation-level computers. However, while the MM5 was nonhydrostatic, it was not an optimal tool to pursue those scales: it was nonconservative, it had low-order numerics (meaning less accurate solutions for finer scales), and it lacked a framework that facilitated the addition of advanced physics or that supported many desirable software attributes [portability, parallelism, extensibility, software layers, and application programming interfaces (APIs)]. Meanwhile, the National Centers for Environmental Prediction (NCEP) had interest in developing a nonhydrostatic model for operational forecasting on finer scales. In this setting, circa 1995, the idea of WRF took shape on the premise that there could be a beneficial synergy in an NWP model shared by the research and operational camps and that would be a next-generation capability moving past known limitations (such as in the MM5). The model could be a common platform for an extensive research community to develop capabilities that operations could readily exploit. Furthermore, an understanding of model performance and needed improvements could be hastened in the crucible of operational use, guiding practical development in return. Thus, the capability would facilitate “research to operations” (R2O) developments while leading to sharpened research and development efforts on identified needs [the operations to research (O2R) path].

Seeing these complementary possibilities, a partnership formed to build WRF. Its original members were NCAR, the National Oceanic and Atmospheric Administration (NOAA) [represented by NCEP and what has become NOAA’s Earth System Research Laboratory (ESRL)], the U.S. Air Force, the Naval Research Laboratory, the University of Oklahoma, and the Federal Aviation Administration. In addition to planning the initial efforts, the partners provided in-kind and other resources to create the system.

The development of the model’s dynamical solver (or dynamical core) and related numerics were the initial foci. Compared to previous models (such as the MM5), what emerged was superior in terms of higher-order numerical accuracy and scalar conservation properties. As these pieces took shape, an innovative software framework ( Michalakes et al. 1999 ) for the model’s dynamics, physics, and input/output (I/O) components was designed. The architecture united the modeling components logically and efficiently while looking ahead to ensure system extensibility, ease of development, and scalable massively-parallel operation. Another development thrust addressed the preprocessors for domain and input preparation while a separate effort tackled data assimilation. The original partners built these capabilities through the efforts of a number of working groups, and the first model release was in December 2000.

The initial physics packages for WRF were ported from the MM5, but community contributions have since delivered a host of schemes providing multiple options for atmospheric physical processes. Other WRF capabilities have been developed through the resources and efforts of interested agencies and universities. Taken as a whole, WRF’s growth reflects a collaborative and communal path: the system’s development has never been solely funded or directed by a single entity.

WRF’s dynamics are represented in its atmospheric fluid flow solvers, or cores. The two original cores had an Eulerian height–based vertical coordinate and a mass-based vertical coordinate ( Klemp et al. 2007 ). WRF, version 2, saw the removal of the height-based version because the limited benefit of having both cores did not justify the extra complexity. In the early 2000s, another solver, the NCEP’s Nonhydrostatic Mesoscale Model (NMM) core ( Janjić et al. 2001 ; Janjić 2003 ) was added to provide an alternative option for NCEP. The two WRF variants were called the Advanced Research version of WRF (ARW; WRF-ARW) and WRF-NMM.

Oversight of the WRF enterprise has evolved over time. Through the early years, the partners coordinated the various efforts. The WRF Executive Oversight Board, which represented each partner organization, handled top-level management. Under this were the Research Applications Board and the Operations Requirements Board, addressing the interests of the research and operational stakeholders, respectively. At the developmental level, various working groups focused on narrower areas, such as numerics, data assimilation, and physics. From the late 2000s, the original top-down direction of WRF has transitioned to a mode of community-driven input, with the responsibility for basic system and community support led by NCAR (see “The WRF community support effort” section).

WRF AND ITS APPLICATIONS.

WRF produces atmospheric simulations. The process has two phases, with the first to configure the model domain(s), ingest the input data, and prepare the initial conditions, and the second to run the forecast model. The forecast model components operate within WRF’s software framework, which handles I/O and parallel-computing communications. WRF is written primarily in Fortran, can be built with a number of compilers, and runs predominately on platforms with UNIX-like operating systems, from laptops to supercomputers. 2 WRF’s architecture has allowed it to be ported to virtually every type of platform in the world’s top 500 supercomputers (see www.top500.org ).

WRF simulations begin with the WRF Preprocessing System (WPS). A series of utilities, the WPS first pulls in geographical information (e.g., topography, land use) to set up the user’s model domains. Next, it ingests, reformats, and interpolates the requisite first-guess atmospheric data (e.g., a global analysis or model forecast) to the user’s domains. Finally, the input fields are put on the model’s vertical levels and lateral boundary conditions are generated. WRF is then ready to run. This is done by the forecast component that contains the dynamical solver and physics packages for atmospheric processes (e.g., microphysics, radiation, planetary boundary layer). 3

Per its design as a research tool, WRF can also be configured to conduct idealized simulations. This capability allows users to study processes in a simplified setting (e.g., reflecting a single sounding or idealized topography) by varying parameters and initial conditions while using limited physics. WRF currently provides 12 idealized scenarios, including baroclinic waves, supercell convection, flow over topography, large-eddy flows, and tropical cyclones. In addition, individual users can readily construct other idealized configurations.

WRF may also be run as a global model on a latitude–longitude grid. While Global WRF was originally built to study planetary atmospheres ( Richardson et al. 2007 ), it has come to be used for terrestrial forecasting, chemistry, and climate applications ( Zhang et al. 2012 ; Karamchandani et al. 2012 ; Jin et al. 2013 ; Hutchinson 2015 ).

Under both real-data and idealized configurations, WRF has been extensively used for research. As WRF is fundamentally a mesoscale model, WRF research applications over the years have run the gamut of topics in mesoscale meteorology: synoptic and mesoscale processes associated with extratropical cyclones, fronts, and jets ( Zhang et al. 2009 ; Schultz and Sienkiewicz 2013 ; Thompson and Eidhammer 2014 ; Rostom and Lin 2015 ; Ganetis and Colle 2015 ; Lu and Deng 2015 ); mesoscale weather events and phenomena ( Powers 2007 ; Shi et al. 2010 ; Brewer et al. 2013 ; Mass et al. 2014 ; Parish et al. 2015 ; DuVivier and Cassano 2013 ); organized convection ( Trier et al. 2006 ; Xu et al. 2015 ; Meng et al. 2012 ; Morrison et al. 2012 ; Akter 2015 ); and hurricanes ( Davis et al. 2008a ; Khain et al. 2010 ; H. Chen et al. 2011 ; Moon and Nolan 2015 ). In recent years WRF’s use for regional climate research has surged (see, e.g., Leung and Qian 2009 ; Done et al. 2015 ; Bruyère et al. 2014 ), and, in this, WRF’s forte is resolving smaller-scale atmospheric and land surface processes better than the global models traditionally used for climate projections.

The WRF Data Assimilation System (WRFDA) is the primary data assimilation system for WRF ( Barker et al. 2004 , 2012 ; Huang et al. 2009 ). It features three-dimensional and four-dimensional variational (3DVAR, 4DVAR) approaches, as well as a hybrid variational–ensemble approach [ensemble transform Kalman filter (ETKF); ETKF-3DVAR; Wang et al. 2008 ; Schwartz et al. 2015c ]. These techniques can assimilate a wide range of direct and indirect observation types, from traditional in situ surface and upper-air data to satellite-based measurements. Table 1 lists the observations that WRFDA can currently ingest. In addition to WRFDA, the Gridpoint Statistical Interpolation analysis system (GSI; Wu et al. 2002 ; Kleist et al. 2009 ) and the Data Assimilation Research Testbed (DART; Anderson et al. 2009 ) are data assimilation capabilities that can be used to prepare WRF initial conditions.

WRFDA Observation Catalog. Data formats used vary with observation but include ASCII, NCEP Prepared BUFR (PrepBUFR), binary, and Hierarchical Data Format (HDF). SYNOP = synoptic. METAR = aviation routine weather report. Pibal = pilot balloon. HIRS = High Resolution Infrared Radiation Sounder. AMSU = Advanced Microwave Sounding Unit. MHS = Microwave Humidity Sounder. SATEM = temperature profile. SSMIS = Special Sensor Microwave Imager/Sounder. DMSP = Defense Meteorological Satellite Program. ATMS = Advanced Technology Microwave Sounder. SNPP = Suomi National Polar-Orbiting Partnership . AIRREP = air report. ACARS = Aircraft Communication, Addressing, and Reporting System. AMDAR = Aircraft Meteorological Data Relay. TAMDAR = Tropospheric Airborne Meteorological Data Reporting. AIRS = Atmospheric Infrared Sounder. GPS = global positioning system. IASI = Infrared Atmospheric Sounding Interferometer. MetOp = Meteorological Operational satellite. SEVIRI = Spinning Enhanced Visible and Infrared Imager. Meteosat = Meteorological Satellite. MWTS = Microwave Temperature Sounder. MWHS = Microwave Humidity Sounder. FY-3 = Fengyun-3. AMSR-2 = Advanced Microwave Scanning Radiometer-2. GCOM = Global Change Observation Mission–Water 1.

New approaches to data assimilation and the impact of new observation types are fertile areas of research using WRFDA (e.g., Jung et al. 2013 ; Xu et al. 2013 ; Romine et al. 2016 ). Additionally, ensemble Kalman filter data assimilation has been used to initialize ensembles of WRF with convection-permitting resolutions (e.g., 3-km grids) for field campaigns [such as the Mesoscale Predictability Experiment (MPEX) in 2013 ( Schwartz et al. 2015a ) and the Deep Convective Clouds and Chemistry (DC3) field campaign of 2012 ( Romine et al. 2014 )], for NOAA’s Hazardous Weather Testbed experiment ( Clark et al. 2012 ), and for ongoing real-time forecasting (e.g., at NCAR; see Schwartz et al. 2015b ). Individual convection-permitting WRF runs have demonstrated the capability to capture the observed evolution and structure of organized convective storms. Figure 3 provides a good example for a long-lived central-U.S. squall line event of July 2015, from the NCAR ensemble. Apart from a timing offset of a few hours, the WRF prediction of over 24 hours in advance matches well the system progression seen in the composite radar analysis.

Envisioned for NWP, WRF is used operationally at governmental centers around the world (see, e.g., Fig. 2 ; Dudhia 2014 ) as well as by private companies. The configurability of high-resolution domains, variety of possible input data, and computational flexibility (particularly in limited-resource settings), along with the ability to leverage model advancements from a global research community, have made it particularly attractive for real-time forecasting.

In the United States, NCEP employs WRF in support of the National Weather Service in a number of systems. WRF (ARW) is run in the Rapid Refresh (RAP) and High-Resolution Rapid Refresh (HRRR) systems developed by the NOAA ESRL and NCEP ( Benjamin et al. 2016 ; Peckham et al. 2016 ). Providing the benefits of an hourly update cycle, the RAP and HRRR systems feature 13- and 3-km grids across North America and the conterminous United States (CONUS), respectively. Their frequent initializations use the GSI data assimilation system, employing a hybrid variational–ensemble approach, as well as digital filter initialization ( Peckham et al. 2016 ). Figure 4 provides an example of an HRRR forecast and verification over the 3-km CONUS domain. HRRR and RAP applications include energy, hydrology, severe weather, aviation weather, and air quality, and are partially linked to several of the WRF specialty systems noted below (see the “Tailored WRF systems and model applications” section). A recently developed example is HRRR-Smoke, which employs the WRF Model coupled with Chemistry (WRF-Chem) in real time with the HRRR and has been developed to simulate the emissions and transport of smoke from wildfires (see https://rapidrefresh.noaa.gov/hrrr/HRRRsmoke/ ).

NCEP also uses WRF in its short-range ensemble forecasting (SREF) system, which includes 13 ARW forecast members (on 16-km grids), and in the High-Resolution Window Forecast System (HIRESW), with domains of 4.2–3.5-km grid spacing for the CONUS, Alaska, Hawaii, Puerto Rico, and Guam. As described below, NCEP also runs a version of the model called Hurricane WRF (HWRF; Tallapragada et al. 2014 ) operationally for tropical cyclone prediction.

From WRF’s inception, the U.S. Air Force has been both a key user of, and contributor to, WRF for real-time NWP. WRF has supported its forecasting needs in theaters across the globe since the mid-2000s. Outside of the United States, Taiwan’s Central Weather Bureau (CWB) runs WRF for operational forecasting, which includes a model configuration for its western Pacific typhoon prediction needs (Typhoon WRF; Hsiao et al. 2012 ). In a very different setting (namely, polar) the Antarctic Mesoscale Prediction System (AMPS; Powers et al. 2012 ) runs WRF over Antarctica to support the U.S. Antarctic Program (USAP). The primary users are the weather forecasters for the USAP, who provide the forecasts for crucial air operations and scientific and logistical activities across the continent.

Apart from its operational real-time user, WRF is run for real-time forecasting by scores of universities worldwide. Its output provides a focus for weather analysis and forecasting by students and faculty, as well as a vehicle for classroom explorations of NWP and course topics. WRF’s portability to workstation and limited-compute environments has facilitated this, and the university application has introduced many students to the practice of NWP.

Real-time WRF also supports science via NWP for field campaigns. In recent years WRF-assisted experiments have included DC3 (2012) ( Barth et al. 2015 ); Tropical Ocean Troposphere Exchange of Reactive Halogen Species and Oxygenated Volatile Organic Compounds (TORERO; 2012) ( Volkamer et al. 2015 ); Studies of Emissions and Atmospheric Composition, Clouds, and Climate Coupling by Regional Surveys (SEAC4RS; 2013) ( Wagner et al. 2015 ); and MPEX (2013) ( Weisman et al. 2015 ), and more recently the Olympic Mountains Experiment (OLYMPEX; http://olympex.atmos.washington.edu/ ) and the O 2 /N 2 Ratio and CO 2 Airborne Southern Ocean Study (ORCAS; www.eol.ucar.edu/field_projects/orcas ). OLYMPEX (November 2015–February 2016) studied precipitation measurements in midlatitude oceanic frontal systems interacting with the coast and mountains, specifically the Olympic Peninsula and Olympic Mountains in Washington and addressed the validation of data from the NASA Global Precipitation Measurement (GPM) program satellite constellation ( www.nasa.gov/mission_pages/GPM/main/index.html ). The experiment included sampling in fronts, extratropical cyclones, and atmospheric rivers, with the University of Washington running real-time WRF down to 4 km in support of the missions of the NASA DC-8 and ER-2 and University of North Dakota Citation aircraft. ORCAS (January–February 2016) has sought to improve the understanding of the physical and biological controls on air–sea exchange of O 2 and CO 2 in the Southern Ocean. Within the framework of AMPS, WRF provided weather guidance for the NCAR High-Performance Instrumented Airborne Platform for Environmental Research (HIAPER; Gulfstream V) aircraft chemical measurement flights over the Southern Ocean and the Antarctic Peninsula.

Private sector real-time use of WRF is significant. Two large concerns that employ WRF for regular NWP services are The Weather Company and Vaisala. The Weather Company ( www.theweathercompany.com ) runs a modified version of WRF under the name Rapid Precision Mesoscale (RPM) to provide a range of services and products, including flight forecasting, which focuses on convection, turbulence, and icing. The Weather Company also has pushed the delivery of WRF-based weather information into the medium of television production and applications for smartphones, tablets, and desktop systems. Vaisala ( www.vaisala.com ), known in the meteorology community for its atmospheric measurement systems and information services, runs WRF for renewable energy efforts, both to assess the potential for wind and solar energy projects and to improve return on existing facility investments. Its services include probabilistic forecasting to address wind energy scheduling issues and to maximize opportunities to boost energy delivery to the electric power grid. As another innovative direction for WRF NWP, Ignitia ( www.ignitia.se ) has targeted agriculture and is running a model configuration in aid of farmers in tropical West Africa, focusing on Ghana ( Smith 2015 ). Here, subscribers receive text messages with WRF forecast information that has been customized for farm decision-making (e.g., deployment of workers or machinery), improving harvest and investment returns.

To address Earth system prediction beyond weather, WRF supports a host of tailored capabilities. Most of these are compiled with WRF and run with it, as opposed to being stand-alone models run offline using WRF output. Thus, the capabilities have strong two-way interaction with the atmospheric component.

Air chemistry

WRF-Chem is a WRF-based in-line atmospheric chemistry model ( Grell et al. 2005 ; Fast et al. 2006 ). It is applied in a wide range of research on air quality and chemistry and has a spectrum of options to handle gas-phase and aqueous chemistry and aerosols. Table 2 lists the main process capabilities and applications of WRF-Chem. WRF-Chem integrates chemistry and dynamics at every time step, which is essential not only for investigating aerosol, weather, and climate interactions, but also for air quality research ( Grell et al. 2004 ; Grell and Baklanov 2011 ; Baklanov et al. 2014 ). WRF-Chem can be used for dispersion forecasts (e.g., volcanic ash, dust, smoke, or other hazardous constituents), as well as the prediction and investigation of air quality (also with aerosol–radiation and aerosol–microphysics interactions) and complex interactions between chemistry, aerosols, and physics (such as aerosol indirect effects and aqueous-phase chemistry).

Main capabilities and applications of WRF-Chem, version 3.8. Descriptions of the parameterizations/schemes for the various processes can be found in the WRF-Chem User’s Guide ( NOAA 2016 ) and in WRF-Chem related publications (see http://ruc.noaa.gov/wrf/wrf-chem/References/WRF-Chem.references.htm ).

The NOAA ESRL provides WRF-Chem community support, including tutorials, and its WRF-Chem help desk provides guidance in model use, access to the code repository, and news on developments. The WRF-Chem development community convenes annually at the WRF Users’ Workshop, and in 2016 a special issue devoted to WRF-Chem was jointly organized between the open-access journals Geoscientific Model Development and Atmospheric Chemistry and Physics (see www.atmos-chem-phys.net/special_issue365_33.html ).

The WRF Hydrological Modeling Extension Package (WRF-Hydro) is a hydrology modeling capability providing both fully coupled two-way interactions with WRF and a stand-alone capability ( Gochis et al. 2015 ). First released to the community in 2013, WRF-Hydro is “multiscale” in that it can represent physical processes such as precipitation, infiltration, snowmelt, hillslope overland flow, and channel flow on varying grids. In addition, it is “multiphysics” in that it offers a number of options for representing hydrologic processes. Figure 5 depicts the WRF-Hydro forecasting elements.

WRF-Hydro includes a real-time streamflow data assimilation system, and outputs include snowpack (snow depth and snow water equivalent), soil moisture, evapotranspiration, standing/ponded water, shallow groundwater, and flow through rivers and reservoirs. The model code has been used in a variety of process and forecasting studies (e.g., David et al. 2009 ; Yucel et al. 2015 ; Senatore et al. 2015 ; Givati et al. 2016 ). Based on a collaboration of NCAR and the NOAA National Water Center, WRF-Hydro was chosen as the framework for the national hydrologic prediction system and has begun operational forecasting.

Fire weather

WRF-Fire ( Coen et al. 2013 ) is a WRF and wildland fire-behavior physics package. The system has two-way coupling with the atmospheric dynamics such that at each time step near-surface WRF winds direct the fire spread rate and direction. In turn, the sensible and latent heat fluxes from the combustion force the atmosphere, producing fire-induced winds. The physics package keeps track of the subgrid-scale interface (the “flaming front”) between burning and unignited areas; calculates the fire spread rate as a function of local wind, slope, and fuel properties (e.g., amount, structure, and moisture); calculates the consumption rate of fuel; and releases heat fluxes into the boundary layer. WRF-Fire has been applied to reproduce fire phenomena ( Simpson et al. 2013 , 2014 ), investigate the effects of fuel moisture content and type ( Coen et al. 2013 ), interpret wildfire case studies ( Peace et al. 2015 ), and predict smoke and air quality ( Kochanski et al. 2015 ). As an illustration of WRF-Fire, Fig. 6 shows output from a simulation of a 2015 Colorado grassland fire on a 200-m grid.

Tropical cyclones

HWRF ( Tallapragada et al. 2014 ) is a version of WRF tailored for operational hurricane forecasting. NCEP, the primary developer, deploys HWRF worldwide for real-time prediction for tropical cyclones for which the National Hurricane Center (NHC), Joint Typhoon Warning Center (JTWC), and Central Pacific Hurricane Center (CPHC) have responsibility. HWRF employs the NMM core and, to better capture air–sea interaction, couples the Princeton Ocean Model (POM; Mellor 2004 ) in the Atlantic and northeast Pacific basins. Plans include the coupling of a surface wave model (NOAA Wavewatch III; Tolman 2009 ) as well.

WRF has been applied extensively for tropical cyclone (TC) research, and the Advanced Hurricane WRF ( Davis et al. 2008b , 2010 ) is an ARW configuration tuned for this purpose. TC studies have exploited WRF’s menu of physics options and moving nest capabilities to investigate storm genesis, intensification, structure, and evolution ( Davis et al. 2008a ; Fierro et al. 2009 ; Fang and Zhang 2010 ; Wang et al. 2010 ; H. Chen et al. 2011 ; Chen and Zhang 2013 ; Miller et al. 2015 ). Furthermore, coupled systems of WRF and ocean and wave models have been built to improve TC forecasts, capture air–sea interactions, and better represent ocean and storm impacts/responses ( Warner et al. 2010 ; B. Liu et al. 2011 ; Gopalakrishnan et al. 2012 ; Kim et al. 2014 ; Liu et al. 2015 ; Chen and Curcic 2016 ).

Also in the vein of better simulating atmosphere–ocean interactions, the Coupled Ocean–Atmosphere–Wave–Sediment Transport (COAWST) system ( Warner et al. 2010 ) was created and applied for hurricane simulations ( Zambon et al. 2014 ). Supported by the Woods Hole Oceanographic Institution, COAWST links WRF to the ocean model Regional Ocean Modeling System (ROMS; Shchepetkin and McWilliams 2005 ), the wave model Simulating Waves Nearshore (SWAN; Booij et al. 1999 ), and the Community Sediment Transport Modeling System ( Warner et al. 2008 ). Similarly, the University of Miami has coupled WRF to the Hybrid Coordinate Ocean Model (HYCOM; Bleck 2002 ) and the University of Miami Wave Model (UMWM; Donelan et al. 2012 ) for hurricane studies ( Chen et al. 2013 ; Chen and Curcic 2016 ).

Urban meteorology

WRF-Urban (F. Chen et al. 2011 ) couples an integrated, multiscale, multiphysics urban modeling system to WRF to represent the impacts of urbanization on regional weather and climate, public health, and water resources. The system integrates localized city morphology datasets and multiple urban-modeling capabilities (e.g., building/structure effect parameterizations) to capture interactions among heat islands, the city canopy boundary layer, and mesoscale atmospheric conditions. WRF-Urban is being applied by more than 100 groups in 25 countries for forecasting weather and air quality for cities (e.g., at the Beijing Institute of Urban Meteorology; Barlage et al. 2016 ), for investigating the impacts of urbanization on regional meteorology and water resources, and for exploring adaptation strategies of urban planners.

Solar and wind energy

WRF-Solar ( Jimenez et al. 2016 ) is a configuration of WRF tailored for solar energy forecasting and applications. Among other modifications, it includes improved solar tracking and parameterizations, allowing either the interpolation of irradiances between expensive radiation scheme calls or the use of a fast radiative transfer code for irradiance calculations. WRF-Solar has improved representations of aerosol–radiation, cloud–aerosol, and cloud–radiation interactions and has been found to yield more accurate aerosol, shortwave irradiance, and solar parameter predictions ( Jimenez et al. 2016 ).

WRF use in wind energy settings (highlighted above as a real-time use) is also growing. Applications in this area are pushing WRF to large-eddy resolutions to predict winds at turbine hub heights (e.g., 80 m AGL; Y. Liu et al. 2011 ). An award-wining WRF-based capability is the NCAR Wind and Solar Power Forecasting System (see Mahoney et al. 2012 ; Haupt and Mahoney 2015 ). 4 This provides for forecasting over wind farms serving the regional utility Xcel Energy, which has the largest wind energy generation capacity in the United States ( Haupt and Mahoney 2015 ). The U.S. Department of Energy’s National Renewable Energy Laboratory has collaborated in the development of the power production algorithms using the WRF data, which target turbine heights and which the system translates into energy generation predictions provided to Xcel.

Large-eddy-scale modeling

The WRF with turbulence-resolving capability based on a large-eddy simulation (LES) approach is called WRF-LES. Through very fine (e.g., 50–100 m) grids it can simulate flows in idealized, canonical atmospheric boundary layers as well as realistic, evolving boundary layers driven by large-scale flows. WRF-LES has been verified in planetary boundary layer (PBL) simulations under a spectrum of stability conditions ( Mirocha et al. 2010 ; Munoz-Esparza et al. 2014 , 2015 ). To develop a multiscale capability that bridges the mesoscale–microscale gap, ongoing developments include a scale-aware PBL scheme for the 100-m to 1-km range, a turbulence-triggering LES approach, and a surface layer parameterization accounting for surface heterogeneity (see, e.g., Mirocha et al. 2014 ; Aitken et al. 2014 ; Xiao et al. 2015 ).

WRF-LES is probing the large-eddy scale ( Moeng et al. 2007 ; Mirocha et al. 2010 ) and aims to represent the scales, energies, and transport associated with atmospheric turbulence. It can simulate flows in hurricane boundary layers ( Zhu 2008 ) and across wind farms (Y. Liu et al. 2011 ), and very high-resolution (e.g., Dx ∼50 m) simulations are being applied to explore both idealized settings ( Kirkil et al. 2012 ) and real-data cases ( Talbot et al. 2012 ). Recent work has significantly improved WRF for large-eddy applications by changing the model’s prognostic formulation of potential temperature ( Xiao et al. 2015 ), with the benefits of avoiding spurious simulated motions and reducing the computational cost of LES runs. Pushing WRF to explore the microscale is helping to shape one of the model’s future roles in atmospheric modeling: addressing scales out of the practical reach of global models while spanning domains beyond which classical large-eddy modeling or direct numerical simulation can be afforded computationally.

Polar environments

Polar WRF ( Hines and Bromwich 2008 ; Bromwich et al. 2009 ; Hines et al. 2011 ; Bromwich et al. 2013 ) provides options to more accurately represent conditions over the high latitudes and ice sheets. These include fractional sea ice representation, adjustments to the thermal/radiative properties of ice and snow surfaces, and specification of sea ice albedo and snow depth on sea ice. While the released version of WRF incorporates most of these modifications, the full Polar WRF code with the latest updates is provided by The Ohio State University ( http://polarmet.osu.edu/PWRF ). Supporting both climate and meteorological studies, the recently prepared Arctic System Reanalysis (ASR) dataset ( Bromwich et al. 2016 ) has been generated from Polar WRF. This regional dataset targeting the Arctic is of higher resolution than the widely used European Centre for Medium-Range Weather Forecasts (ECMWF) interim reanalysis (ERA-Interim) and has shown improvements over it (Bromwich et al. 2015).

WRF is used at universities around the world as a vehicle for teaching and thesis work. Based on a survey of professors at University Corporation for Atmospheric Research (UCAR) member institutions, the ways in which the system is used include instruction on NWP, numerical methods, and atmospheric processes; research projects; real-time runs for forecasting and weather classes; and model output for lecture and class materials. For studies, both graduate and undergraduate students can readily apply WRF to explore a spectrum of atmospheric research problems, using its idealized, real-data, or specialty capabilities. Furthermore, they can quickly begin to address scientific questions without having to reinvent and build a complex model as a study tool. That WRF can readily be installed and run on common computer hardware (i.e., workstations, clusters, and laptops) has been a big factor in its use in education.

THE WRF COMMUNITY SUPPORT EFFORT.

In the early years, the original WRF partners directed the development and support of WRF. As the system became established, the broader community increasingly undertook development. Model support, however, has remained centralized in NCAR, and today that effort includes user assistance, system oversight, and integration of code contributions.

Part of NCAR’s mission is to support the university community’s atmospheric research through facilities. Thus, as it had originally supported the MM5, NCAR transitioned that role to WRF, and NCAR currently manages and oversees the system. Of the original WRF management structure, the Research Applications Board (RAB) and certain working groups continue. The RAB, composed of scientists from the community, convenes during the annual WRF Users’ Workshop and considers WRF scientific issues and research community interests.

NCAR, and specifically its Mesoscale and Microscale Meteorology (MMM) Laboratory, is responsible for the management of the WRF code, for the oversight of releases, and for providing community support. To coordinate code and release management with community participation, the WRF Developers’ Committee and the WRF Release Committee were established. The Developers’ Committee (DC) has responsibility for maintaining the WRF system software by approving and overseeing code contributions, code testing, and repository upkeep. Members of the DC are active in creating code and in maintaining the WRF system, and they shepherd new code into the WRF repository by being liaisons with external contributors.

The WRF Release Committee (RC) oversees the model major release process. It sets the release schedule, compiles a list of features proposed by developers, and reviews the progress of candidate items. The committee makeup reflects the active areas of WRF code development (e.g., software, physics, data assimilation, chemistry), and its members serve as points of contact with contributors in these areas. Major releases are issued annually (April), with typically subsequent minor (primarily bug fix) releases by the following September.

Since WRF’s initial development, new components and capabilities have come from directed funding by agencies and from contributions from those in the research and operational communities. Anyone may offer code to WRF: submissions are not restricted to prescribed groups. Among the requirements for contributions, however, developers are responsible for the testing of their code and for providing the results to the DC (see NCAR 2016 ).

The WRF Physics Review Panel reviews new physics packages that are submitted by developers for inclusion in the model repository. Functioning like a journal editor, the panel communicates with the prospective contributors and taps relevant experts from the scientific community to anonymously review the proposed packages. The review materials required from developers include testing results and code documentation. The panel considers the reviewers’ input, with criteria that include scientific soundness, novelty of approach, and projected community interest, and makes a recommendation to the Developers’ Committee on whether the package should be accepted.

Furnishing user support to the worldwide WRF community is a major effort. For the core WRF system (WRF and WRFDA), the MMM Laboratory provides this mainly under its base budget from the National Science Foundation. For WRF specialty systems, the primary development groups provide support (e.g., NOAA/ESRL for WRF-Chem). Fundamental to core WRF support is user assistance, provided chiefly through the wrfhelp service. Wrfhelp receives from 350 to 475 inquiries a month, ranging from questions on input data, to model configuration and compilation, to run-time problems. 5 Model support also includes maintaining the user pages and downloadable materials, such as the WRF code, utilities, and documentation, and an important function is working with developers to integrate code contributions for releases. Users can also get assistance informally from other, experienced users through an online WRF forum ( http://forum.wrfforum.com ).

Tutorials have been a pillar of community support. The primary tutorials are hosted by NCAR in January and July, are a week long, and consist of lectures and practice sessions. Class sizes are typically 60. There are also regular tutorials for WRFDA and WRF-Chem. The NCAR support team also periodically delivers tutorials abroad.

WRF support includes organizing and hosting the annual WRF Users’ Workshop. Held at NCAR (Boulder, Colorado) each June, the workshop brings together the WRF community to discuss model developments, results, and issues. Popular workshop sessions include lectures on a selected modeling topic (e.g., cumulus parameterizations, data assimilation), group discussions on the main WRF-related areas (e.g., physics, software), and mini tutorials on model-related utilities (e.g., visualization tools). The workshop averages about 200 registrants per year.

Though mature, WRF continues to advance and to effectively serve the needs of its user groups through the creative developments and applications of the community. Sketched below are future directions for WRF. It is emphasized that the examples are not exhaustive and that the varied interests of the user community will ultimately draw the picture.

While we have presented many positives of the model, WRF is not without issues and challenges. As one example, although WRF computes efficiently in parallel, it can exhibit relatively slow I/O performance in certain massively parallel environments. Currently, to reduce I/O time WRF does have an option to produce separate output files for each processor used. However, in settings with very high CPU counts, the number of resultant files can become a problem. Thus, to improve WRF’s parallel outputting, a future release will offer parallel I/O (PIO), a library for handling output that can scale in parallel with very large WRF domains.

An example of a challenge to WRF is the processing of code contributions. One issue here is ensuring the value of new physics packages, in light of the accumulation of options in the model. In response to this, the WRF Physics Review Panel was set up. Second, the testing and integration of new code require the resources of the WRF community support team. To reduce this demand, as well as to improve the level of submitted code, more stringent requirements on contributors for code testing and documentation have been established. As a related issue, another notable challenge to WRF is the pressure on the community support effort, in light of some erosion of the funding for this work. The community should note, however, that the WRF effort will continue to provide basic support (e.g., user help) and will prioritize that over other activities as necessary.

Considering the future demand for WRF, as long as it can more effectively meet users’ needs as an accurate, extensible, and scalable regional modeling system, it should continue to be widely employed. WRF is an established, understood system with a long track record. It has physics and specialty systems for capturing a range of Earth system processes, and the needs for these types of capabilities will continue. Furthermore, WRF can focus on regional domains with resolutions beyond what most users can attain using global models. One future role for WRF is thus in operating in the spectral range between the applied scales of global and LES models: the space from the convective to large-eddy scales.

In terms of WRF development, one direction is that of connected physics packages. With this, one scheme sends prognostic quantities to another, instead of the latter relying on its own estimates (e.g., from climatologies). An example is the coupling of the Thompson microphysics scheme and Rapid Radiative Transfer Model for general circulation models (RRTMG) radiation scheme, as described in Thompson et al. (2016) . Whereas the radiation scheme would assume sizes of water and ice species, here the microphysics scheme supplies the predicted values of effective radii of cloud water, ice, and snow ( Thompson and Eidhammer 2014 ; Thompson et al. 2016 ). Similar work is addressing aerosol information for microphysics and radiation schemes, and cumulus and radiation scheme coupling has also been done ( Alapaty et al. 2012 ). Leveraging the prognostic output from physics packages and increasing the consistency of assumptions and calculations within the model are benefits from the connected physics approach.

Related to coupled physics is the idea of defining physics suites in WRF. Physics suites are combinations of schemes (e.g., microphysics, radiation, cumulus, PBL) that are run together. In WRF these can reflect schemes that have been coupled or package combinations that are known, or may be tuned, to perform well together. Motivations for designating suites are to improve the understanding of model behavior and ultimately improve model performance: concentrated use of a suite could focus attention on a scheme set, accelerate improvements within that set, and yield a better-documented model configuration. Having suites known for their suitability for specific grid sizes and applications (e.g., regional climate or hurricane prediction) can guide users in configuring the model, and the more consistent use and verification of a given combination can promote its further development.

The implementation of scale-aware physics schemes is another goal for WRF. These are packages that aim to perform accurately over a range of horizontal grid spacings without the problems arising from the breakdown of assumptions tied to specific scales (see, e.g., Molinari and Dudek 1992 ). Such schemes would clearly benefit applications with nesting that enters the “gray zone” between parameterized and explicit scales of convection (see, e.g., Wyngaard 2004 ; Yu and Lee 2010 ; Hong and Dudhia 2012 ; Shin and Hong 2013 ; Ching et al. 2014 ). Ideally, a scale-aware physics scheme would run accurately from hydrostatic (e.g., ≥10 km), to convection-permitting (e.g., 1–3 km), to large-eddy (e.g., tens/hundreds of meters) scales. Recently, Shin and Hong (2015) presented an algorithm for vertical transport in convective boundary layers aimed to bridge the subgrid- to resolved-scale gap, and such approaches may aid in the development of scale-aware schemes for WRF.

The future will see continued work to advance WRF computational performance. First, efforts are underway to speed up popular physics packages. For example, software engineers at the University of Wisconsin have worked on the Goddard longwave and shortwave radiation packages, the Yonsei University (YSU) PBL scheme, and the Thompson microphysics scheme, with plans to target other schemes as well. Their recoding improvements have been based on analyses from benchmarking and instrumentation programs. Preliminary results are showing substantive speedups (e.g., 2–4 times) in the selected schemes while maintaining bit-for-bit identical results ( Mielikainen et al. 2014 ; Huang et al. 2014 ).

Second, to improve overall code performance and to more effectively use high-performance computing (HPC) platforms, Open Multiprocessing (OpenMP) compute environments for WRF will be exploited. This is being done through hybrid parallelism, in which there is shared-memory parallel operation (i.e., using OpenMP) within a compute node and a distributed-memory parallel approach [using message passing interface (MPI)] for communications between nodes. Testing in WRF has shown that hybrid parallelism can be better than either pure MPI or OpenMP. In addition, the WRF framework’s emphasis on MPI tasks and OpenMP threads aligns well with next-generation computers adopting the Many Integrated Core (MIC) approach ( Satish et al. 2010 ). This strategy uses large numbers of lower-performance processors on a single chip to increase hardware parallelism, computations per watt, and overall performance.

As suggested above, WRF will continue to adapt to, and exploit, new computing trends and designs. Supercomputing is moving toward architectures with legions of cheaper, lower-clock-speed, and more energy-efficient processors. For instance, the last three main community machines at NCAR have featured increased processor counts and decreased processor clock speeds: Bluefire (2008–13) with 4,064 chips, each 4.7 GHz; Yellowstone (2012–17) with 72,576 chips, each 2.6 GHz; and Cheyenne (2017 and beyond) with 145,152 chips, each with 2.3 GHz. WRF runs well on these architectures. However, an opposite paradigm to that of supercomputer mainframes is also emerging: cloud computing. This is the use of the Internet to enlist a large number of remote servers for job processing, and WRF can take advantage of it. As an example, Amazon Web Services ( https://aws.amazon.com ) offers software and tools for users to create a compute cluster for doing WRF simulations via the cloud. Cloud computing for WRF promises users the ability to do modeling requiring compute resources otherwise beyond their hardware means.

One final aspect of the future of WRF is its relationship to other models, in particular global models and the Model for Prediction Across Scales (MPAS) project ( Skamarock et al. 2012 ). Variable-resolution global models, such as MPAS, may provide clear benefits over regional models that rely on imposed lateral boundary conditions derived from those larger-scale models (see, e.g., Park et al. 2014 ). As we expect, however, that WRF will continue to be a preferred modeling platform for increasingly higher-resolution, shorter-time-scale applications (LES, wind energy, urban meteorology, cloud and storm dynamics, etc.) for the foreseeable future, there will be substantial benefits in the sharing of components and capabilities across system types.

The likely areas of interface for WRF will be physics, couplers, and postprocessing. The ability of WRF to share specific physics scheme sets with other models (e.g., MPAS) would support the testing of packages across scales and comparisons of model platforms. It would also allow WRF to use the same physics as the model supplying its first-guess field and boundary conditions, giving a measure of consistency of initial conditions (ICs)/boundary conditions (BCs) and the simulation. As an example of WRF’s move toward shared physics, the WRF and MPAS efforts are currently considering a common physics repository.

A specific coupler or coupling approach is another capability that WRF could adopt in common with other atmospheric modeling systems (CESM, MPAS, etc.). This would allow WRF to join readily with other types of environmental models (e.g., ocean, wave), using the same developed and supported connectivity, and it could allow for coupled WRF systems incorporating the same Earth process components as global systems. Similarly, a postprocessing capability that WRF shared with MPAS could provide an improved tool for output analysis (e.g., diagnostics and graphics) and the multimodel user interest in the common capability could promote its development and facilitate its support.

In retrospect, WRF’s significance for meteorology and atmospheric modeling in large part rests on the fact that over the years it has not simply supported—but stimulated—a productive and evolving community by providing solid common ground on which to pursue ideas and build on results. WRF has pioneered in, and expanded the boundaries of, high-resolution and coupled atmospheric prediction. Over its tenure, it has supported and served research, education, public needs, and private interests.

Looking forward, WRF will continue to enable the research and NWP communities as a powerful resource. Though computing and modeling advancements are allowing global models to address the mesoscale with higher resolution, WRF will in turn be able to capture regional and smaller scales more effectively with even finer grids, reaching down to large-eddy grid spacing. WRF will continue to be an effective platform for developing improved representations of physical processes and for serving as the foundation of an array of tailored and coupled systems for integrated Earth system modeling. WRF is truly a community model, and its populous, diverse user base will keep driving it to meet an evolving range of scientific, operational, educational, and commercial needs. It is envisioned that the community’s innovative applications and contributions will advance a vital WRF Model for years to come.

ACKNOWLEDGMENTS

On behalf of itself and the community, NCAR thanks the National Science Foundation for its funding of WRF community support over the years.

This is the accumulated number of registrants. The number of those actively running the model cannot be known. In comparison, the prominent community climate model Community Earth System Model (CESM; Hurrell et al. 2013 ) currently has about 6,000 registered users, with new registrations for the 2010–15 period averaging over 700 per year.

Capabilities also exist for WRF to be run on Windows operating systems through software containers that package the model application for the given platform.

For full details on the components and operation of WRF, readers may consult the ARW ( http://www2.mmm.ucar.edu/wrf/users/docs/user_guide_V3/contents.html ) ( NCAR 2017 ) and NMM ( www.dtcenter.org/wrf-nmm/users/docs/user_guide/V3/users_guide_nmm_chap1-7.pdf ) ( DTC/NCEP 2014 ) users' guides.

This system was the recipient of the 2014 Colorado Governor’s Office Colorado Leveraging Assets for Better Science (CO-LABS) Sustainability Award.

Users may contact wrfhelp via e-mail to [email protected] .

AMS Publications

Get Involved with AMS

Affiliate sites.

Email : [email protected]

Phone : 617-227-2425

Fax : 617-742-8718

Headquarters:

45 Beacon Street

Boston, MA 02108-3693

1200 New York Ave NW

Washington, DC 200005-3928

• Privacy Policy & Disclaimer
• Get Adobe Acrobat Reader
• © 2024 American Meteorological Society
• [185.148.24.167]
• 185.148.24.167

Character limit 500 /500

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Journal Proposal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Machine learning in weather prediction and climate analyses—applications and perspectives.

1. Introduction

2. materials and methods, 3.1. numerical weather prediction, 3.1.1. photovoltaic and wind energy, 3.1.2. atmospheric physics and processes, 3.2. climate, 3.2.1. parametrizations, 3.2.2. extreme events, 3.2.3. climate change, 4. discussion and conclusions, supplementary materials, author contributions, institutional review board statement, informed consent statement, data availability statement, conflicts of interest, abbreviations.

• Fradkov, A.L. Early History of Machine Learning. IFAC-PapersOnLine 2020 , 53 , 1385–1390. [ Google Scholar ] [ CrossRef ]
• Mahesh, B. Machine Learning Algorithms—A Review ; International Journal of Science and Research: Raipur, India, 2019. [ Google Scholar ] [ CrossRef ]
• Dhall, D.; Kaur, R.; Juneja, M. Machine Learning: A Review of the Algorithms and Its Applications. In Lecture Notes in Electrical Engineering, Proceedings of the ICRIC, Jammu, India, 8–9 March 2019 ; Singh, P.K., Kar, A.K., Singh, Y., Kolekar, M.H., Tanwar, S., Eds.; Springer International Publishing: Cham, Switzerland, 2020; pp. 47–63. [ Google Scholar ] [ CrossRef ]
• Singh, A.; Thakur, N.; Sharma, A. A Review of Supervised Machine Learning Algorithms. In Proceedings of the 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), New Delhi, India, 16–18 March 2016; pp. 1310–1315. [ Google Scholar ]
• Iverson, L.R.; Prasad, A.M.; Liaw, A. New Machine Learning Tools for Predictive Vegetation Mapping after Climate Change: Bagging and Random Forest Perform Better than Regression Tree Analysis. In Landscape Ecology of Trees and Forests, Proceedings of the 12th Annual IALE(UK) Conference, Cirencester, UK, 21–24 June 2004 ; Smithers, R., Ed.; International Association for Landscape Ecology: Piestany, Slovakia, 2004; pp. 317–320. [ Google Scholar ]
• Huang, L.; Kang, J.; Wan, M.; Fang, L.; Zhang, C.; Zeng, Z. Solar Radiation Prediction Using Different Machine Learning Algorithms and Implications for Extreme Climate Events. Front. Earth Sci. 2021 , 9 , 202. [ Google Scholar ] [ CrossRef ]
• Abbot, J.; Marohasy, J. The Application of Machine Learning for Evaluating Anthropogenic versus Natural Climate Change. GeoResJ 2017 , 14 , 36–46. [ Google Scholar ] [ CrossRef ]
• Davenport, F.V.; Diffenbaugh, N.S. Using Machine Learning to Analyze Physical Causes of Climate Change: A Case Study of U.S. Midwest Extreme Precipitation. Geophys. Res. Lett. 2021 , 48 , e2021GL093787. [ Google Scholar ] [ CrossRef ]
• Tripathi, S.; Srinivas, V.V.; Nanjundiah, R.S. Downscaling of Precipitation for Climate Change Scenarios: A Support Vector Machine Approach. J. Hydrol. 2006 , 330 , 621–640. [ Google Scholar ] [ CrossRef ]
• Fang, Y.; Chen, H.; Lin, Y.; Zhao, C.; Lin, Y.; Zhou, F. Classification of Northeast China Cold Vortex Activity Paths in Early Summer Based on K-Means Clustering and Their Climate Impact. Adv. Atmos. Sci. 2021 , 38 , 400–412. [ Google Scholar ] [ CrossRef ]
• Zong, L.; Yang, Y.; Gao, M.; Wang, H.; Wang, P.; Zhang, H.; Wang, L.; Ning, G.; Liu, C.; Li, Y.; et al. Large-Scale Synoptic Drivers of Co-Occurring Summertime Ozone and PM2.5 Pollution in Eastern China. Atmos. Chem. Phys. 2021 , 21 , 9105–9124. [ Google Scholar ] [ CrossRef ]
• Silge, J.; Robinson, D. tidytext: Text Mining and Analysis Using Tidy Data Principles in R. J. Open Sour. Softw. 2016 , 1 , 37. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• R Core Team. R: A Language and Environment for Statistical Computing ; R Foundation for Statistical Computing: Vienna, Austria, 2017; Available online: https://www.R-project.org/ (accessed on 17 December 2021).
• Kosovic, B.; Haupt, S.E.; Adriaansen, D.; Alessandrini, S.; Wiener, G.; Delle Monache, L.; Liu, Y.; Linden, S.; Jensen, T.; Cheng, W.; et al. A Comprehensive Wind Power Forecasting System Integrating Artificial Intelligence and Numerical Weather Prediction. Energies 2020 , 13 , 1372. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Dong, W.; Yang, Q.; Fang, X. Multi-Step Ahead Wind Power Generation Prediction Based on Hybrid Machine Learning Techniques. Energies 2018 , 11 , 1975. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Barque, M.; Martin, S.; Vianin, J.E.N.; Genoud, D.; Wannier, D. Improving Wind Power Prediction with Retraining Machine Learning Algorithms. In Proceedings of the 2018 International Workshop on Big Data and Information Security (IWBIS), Jakarta, Indonesia, 12–13 May 2018; pp. 43–48. [ Google Scholar ] [ CrossRef ]
• Ma, Y.-J.; Zhai, M.-Y. A Dual-Step Integrated Machine Learning Model for 24h-Ahead Wind Energy Generation Prediction Based on Actual Measurement Data and Environmental Factors. Appl. Sci. 2019 , 9 , 2125. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Wu, Y.-K.; Su, P.-E.; Wu, T.-Y.; Hong, J.-S.; Hassan, M.Y. Probabilistic Wind Power Forecasting Using Weather Ensemble Models. IEEE Trans. Ind. Appl. 2018 , 54 , 5609–5620. [ Google Scholar ] [ CrossRef ]
• Kim, D.; Hur, J. Short-Term Probabilistic Forecasting of Wind Energy Resources Using the Enhanced Ensemble Method. Energy 2018 , 157 , 211–226. [ Google Scholar ] [ CrossRef ]
• Donadio, L.; Fang, J.; Porté-Agel, F. Numerical Weather Prediction and Artificial Neural Network Coupling for Wind Energy Forecast. Energies 2021 , 14 , 338. [ Google Scholar ] [ CrossRef ]
• Bochenek, B.; Jurasz, J.; Jaczewski, A.; Stachura, G.; Sekuła, P.; Strzyżewski, T.; Wdowikowski, M.; Figurski, M. Day-Ahead Wind Power Forecasting in Poland Based on Numerical Weather Prediction. Energies 2021 , 14 , 2164. [ Google Scholar ] [ CrossRef ]
• Theocharides, S.; Venizelou, V.; Makrides, G.; Georghiou, G.E. Day-Ahead Forecasting of Solar Power Output from Photovoltaic Systems Utilising Gradient Boosting Machines. In Proceedings of the 2018 IEEE 7th World Conference on Photovoltaic Energy Conversion (WCPEC) (A Joint Conference of 45th IEEE PVSC, 28th PVSEC 34th EU PVSEC), Waikoloa, HI, USA, 10–15 June 2018; pp. 2371–2375. [ Google Scholar ] [ CrossRef ]
• Mellit, A. An Overview on the Application of Machine Learning and Deep Learning for Photovoltaic Output Power Forecasting. In Lecture Notes in Electrical Engineering, Proceedings of the 2nd International Conference on Electronic Engineering and Renewable Energy Systems, Saidia, Morocco, 13–15 April 2020 ; Hajji, B., Mellit, A., Marco, T.G., Rabhi, A., Launay, J., Naimi, S.E., Eds.; Springer: Singapore, 2021; pp. 55–68. [ Google Scholar ] [ CrossRef ]
• Mathe, J.; Miolane, N.; Sebastien, N.; Lequeux, J. PVNet: A LRCN Architecture for Spatio-Temporal Photovoltaic PowerForecasting from Numerical Weather Prediction. arXiv 2020 , arXiv:1902.01453. [ Google Scholar ]
• Gómez, J.L.; Martínez, A.O.; Pastoriza, F.T.; Garrido, L.F.; Álvarez, E.G.; García, J.A.O. Photovoltaic Power Prediction Using Artificial Neural Networks and Numerical Weather Data. Sustainability 2020 , 12 , 10295. [ Google Scholar ] [ CrossRef ]
• Theocharides, S.; Makrides, G.; Livera, A.; Theristis, M.; Kaimakis, P.; Georghiou, G.E. Day-Ahead Photovoltaic Power Production Forecasting Methodology Based on Machine Learning and Statistical Post-Processing. Appl. Energy 2020 , 268 , 115023. [ Google Scholar ] [ CrossRef ]
• Kirkwood, C.; Economou, T.; Odbert, H.; Pugeault, N. A Framework for Probabilistic Weather Forecast Post-Processing across Models and Lead Times Using Machine Learning. Philos. Trans. Royal Soc. A Math. Phys. Eng. Sci. 2021 , 379 , 20200099. [ Google Scholar ] [ CrossRef ]
• Cho, D.; Yoo, C.; Im, J.; Cha, D.-H. Comparative Assessment of Various Machine Learning-Based Bias Correction Methods for Numerical Weather Prediction Model Forecasts of Extreme Air Temperatures in Urban Areas. Earth Sp. Sci. 2020 , 7 , e2019EA000740. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Belochitski, A.; Krasnopolsky, V. Stable Emulation of an Entire Suite of Model Physics in a State-of-the-Art GCM Using a Neural Network. arXiv 2021 , arXiv:2103.07028. [ Google Scholar ]
• Walsh, E.; Bessardon, G.; Gleeson, E.; Ulmas, P. Using Machine Learning to Produce a Very High Resolution Land-Cover Map for Ireland. Adv. Sci. Res. 2021 , 18 , 65–87. [ Google Scholar ] [ CrossRef ]
• Seifert, A.; Rasp, S. Potential and Limitations of Machine Learning for Modeling Warm-Rain Cloud Microphysical Processes. J. Adv. Model. Earth Syst. 2020 , 12 , e2020MS002301. [ Google Scholar ] [ CrossRef ]
• Wang, Y.; Zeng, X.; Decker, J. A GPU-Accelerated Radiation Transfer Model Using the Lattice Boltzmann Method. Atmosphere 2021 , 12 , 1316. [ Google Scholar ] [ CrossRef ]
• Bodini, N.; Lundquist, J.K.; Optis, M. Can Machine Learning Improve the Model Representation of Turbulent Kinetic Energy Dissipation Rate in the Boundary Layer for Complex Terrain? Geosci. Model Dev. 2020 , 13 , 4271–4285. [ Google Scholar ] [ CrossRef ]
• Harder, P.; Watson-Parris, D.; Strassel, D.; Gauger, N.; Stier, P.; Keuper, J. Emulating Aerosol Microphysics with Machine Learning. arXiv 2021 , arXiv:2109.10593. [ Google Scholar ]
• Moosavi, A.; Rao, V.; Sandu, A. A Learning Based Approach for Uncertainty Analysis in Numerical Weather Prediction Models. arXiv 2018 , arXiv:1802.08055. [ Google Scholar ]
• Moosavi, A.; Rao, V.; Sandu, A. Machine Learning Based Algorithms for Uncertainty Quantification in Numerical Weather Prediction Models. J. Comput. Sci. 2021 , 50 , 101295. [ Google Scholar ] [ CrossRef ]
• O’Gorman, P.A.; Dwyer, J.G. Using Machine Learning to Parameterize Moist Convection: Potential for Modeling of Climate, Climate Change, and Extreme Events. J. Adv. Model. Earth Syst. 2018 , 10 , 2548–2563. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Yuval, J.; O’Gorman, P.A. Stable Machine-Learning Parameterization of Subgrid Processes for Climate Modeling at a Range of Resolutions. Nat. Commun. 2020 , 11 , 3295. [ Google Scholar ] [ CrossRef ]
• Pielke, R.A. Chapter 8—New Parameterization Approaches. In Mesoscale Meteorological Modeling ; International Geophysics; Pielke, R.A., Ed.; Academic Press: Cambridge, MA, USA, 2013; Volume 98, pp. 225–235. [ Google Scholar ] [ CrossRef ]
• Brajard, J.; Carrassi, A.; Bocquet, M.; Bertino, L. Combining Data Assimilation and Machine Learning to Infer Unresolved Scale Parametrization. Philos. Trans. Royal Soc. A Math. Phys. Eng. Sci. 2021 , 379 , 20200086. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Watt-Meyer, O.; Brenowitz, N.D.; Clark, S.K.; Henn, B.; Kwa, A.; McGibbon, J.; Perkins, W.A.; Bretherton, C.S. Correcting Weather and Climate Models by Machine Learning Nudged Historical Simulations. Geophys. Res. Lett. 2021 , 48 , e2021GL092555. [ Google Scholar ] [ CrossRef ]
• Couvreux, F.; Hourdin, F.; Williamson, D.; Roehrig, R.; Volodina, V.; Villefranque, N.; Rio, C.; Audouin, O.; Salter, J.; Bazile, E.; et al. Process-Based Climate Model Development Harnessing Machine Learning: I. A Calibration Tool for Parameterization Improvement. J. Adv. Model. Earth Syst. 2021 , 13 , e2020MS002217. [ Google Scholar ] [ CrossRef ]
• Nowack, P.; Braesicke, P.; Haigh, J.; Abraham, N.L.; Pyle, J.; Voulgarakis, A. Using Machine Learning to Build Temperature-Based Ozone Parameterizations for Climate Sensitivity Simulations. Environ. Res. Lett. 2018 , 13 , 104016. [ Google Scholar ] [ CrossRef ]
• Alfonso, L.; Zamora, J.M. A Two-Moment Machine Learning Parameterization of the Autoconversion Process. Atmos. Res. 2021 , 249 , 105269. [ Google Scholar ] [ CrossRef ]
• Ukkonen, P.; Pincus, R.; Hogan, R.J.; Pagh Nielsen, K.; Kaas, E. Accelerating Radiation Computations for Dynamical Models With Targeted Machine Learning and Code Optimization. J. Adv. Model. Earth Syst. 2020 , 12 , e2020MS002226. [ Google Scholar ] [ CrossRef ]
• Beucler, T.; Rasp, S.; Pritchard, M.; Gentine, P. Achieving Conservation of Energy in Neural Network Emulators for Climate Modeling. arXiv 2019 , arXiv:1906.06622. [ Google Scholar ]
• Ramadhan, A.; Marshall, J.; Souza, A.; Wagner, G.L.; Ponnapati, M.; Rackauckas, C. Capturing Missing Physics in Climate Model Parameterizations Using Neural Differential Equations. arXiv 2020 , arXiv:2010.12559. [ Google Scholar ]
• Beucler, T.; Pritchard, M.; Yuval, J.; Gupta, A.; Peng, L.; Rasp, S.; Ahmed, F.; O’Gorman, P.A.; Neelin, J.D.; Lutsko, N.J.; et al. Climate-Invariant Machine Learning. arXiv 2021 , arXiv:2112.08440. [ Google Scholar ]
• Krasnopolsky, V.M.; Fox-Rabinovitz, M.S. Complex Hybrid Models Combining Deterministic and Machine Learning Components for Numerical Climate Modeling and Weather Prediction. Neural Netw. 2006 , 19 , 122–134. [ Google Scholar ] [ CrossRef ]
• Gentine, P.; Pritchard, M.; Rasp, S.; Reinaudi, G.; Yacalis, G. Could Machine Learning Break the Convection Parameterization Deadlock? Geophy. Res. Lett. 2018 , 45 , 5742–5751. [ Google Scholar ] [ CrossRef ]
• Rasp, S. Coupled Online Learning as a Way to Tackle Instabilities and Biases in Neural Network Parameterizations. arXiv 2020 , arXiv:1907.01351. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Brenowitz, N.D.; Beucler, T.; Pritchard, M.; Bretherton, C.S. Interpreting and Stabilizing Machine-Learning Parametrizations of Convection. J. Atmos. Sci. 2020 , 77 , 4357–4375. [ Google Scholar ] [ CrossRef ]
• Brenowitz, N.D.; Henn, B.; McGibbon, J.; Clark, S.K.; Kwa, A.; Perkins, W.A.; Watt-Meyer, O.; Bretherton, C.S. Machine Learning Climate Model Dynamics: Offline versus Online Performance. arXiv 2020 , arXiv:2011.03081. [ Google Scholar ]
• Gagne, D.J., II; Christensen, H.M.; Subramanian, A.C.; Monahan, A.H. Machine Learning for Stochastic Parameterization: Generative Adversarial Networks in the Lorenz ’96 Model. J. Adv. Model. Earth Syst. 2020 , 12 , e2019MS001896. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Gettelman, A.; Gagne, D.J.; Chen, C.-C.; Christensen, M.W.; Lebo, Z.J.; Morrison, H.; Gantos, G. Machine Learning the Warm Rain Process. J. Adv. Model. Earth Syst. 2021 , 13 , e2020MS002268. [ Google Scholar ] [ CrossRef ]
• Silva, S.J.; Ma, P.-L.; Hardin, J.C.; Rothenberg, D. Physically Regularized Machine Learning Emulators of Aerosol Activation. Geosci. Model Dev. 2021 , 14 , 3067–3077. [ Google Scholar ] [ CrossRef ]
• Pielke, R., Sr.; Stokowski, D.; Wang, J.-W.; Vukicevic, T.; Leoncini, G.; Matsui, T.; Castro, C.L.; Niyogi, D.; Kishtawal, C.M.; Biazar, A.; et al. Satellite-Based Model Parameterization of Diabatic Heating. Eos Trans. Am. Geophys. Un. 2007 , 88 , 96–97. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Wang, X.; Han, Y.; Xue, W.; Yang, G.; Zhang, G.J. Stable Climate Simulations Using a Realistic GCM with Neural Network Parameterizations for Atmospheric Moist Physics and Radiation Processes. Geosci. Model Dev. Discuss. 2021 , 1–35. [ Google Scholar ] [ CrossRef ]
• Yuval, J.; O’Gorman, P.A.; Hill, C.N. Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision. Geophys. Res. Lett. 2021 , 48 , e2020GL091363. [ Google Scholar ] [ CrossRef ]
• Anochi, J.A.; de Almeida, V.A.; de Campos Velho, H.F. Machine Learning for Climate Precipitation Prediction Modeling over South America. Remote Sens. 2021 , 13 , 2468. [ Google Scholar ] [ CrossRef ]
• Anderson, G.J.; Lucas, D.D. Machine Learning Predictions of a Multiresolution Climate Model Ensemble. Geophys. Res. Lett. 2018 , 45 , 4273–4280. [ Google Scholar ] [ CrossRef ]
• Mansfield, L.A.; Nowack, P.J.; Kasoar, M.; Everitt, R.G.; Collins, W.J.; Voulgarakis, A. Predicting Global Patterns of Long-Term Climate Change from Short-Term Simulations Using Machine Learning. NPJ Clim. Atmos. Sci. 2020 , 3 , 1–9. [ Google Scholar ] [ CrossRef ]
• Hourdin, F.; Williamson, D.; Rio, C.; Couvreux, F.; Roehrig, R.; Villefranque, N.; Musat, I.; Fairhead, L.; Diallo, F.B.; Volodina, V. Process-Based Climate Model Development Harnessing Machine Learning: II. Model Calibration From Single Column to Global. J. Adv. Model. Earth Syst. 2021 , 13 , e2020MS002225. [ Google Scholar ] [ CrossRef ]
• Bochenek, B.; Ustrnul, Z.; Wypych, A.; Kubacka, D. Machine Learning-Based Front Detection in Central Europe. Atmosphere 2021 , 12 , 1312. [ Google Scholar ] [ CrossRef ]
• Liu, Y.; Racah, E.; Prabhat; Correa, J.; Khosrowshahi, A.; Lavers, D.; Kunkel, K.; Wehner, M.; Collins, W. Application of Deep Convolutional Neural Networks for Detecting Extreme Weather in Climate Datasets. arXiv 2016 , arXiv:1605.01156. [ Google Scholar ]
• Biard, J.C.; Kunkel, K.E. Automated Detection of Weather Fronts Using a Deep Learning Neural Network. Adv. Stat. Climatol. Meteorol. Oceanogr. 2019 , 5 , 147–160. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Niebler, S.; Miltenberger, A.; Schmidt, B.; Spichtinger, P. Automated Detection and Classification of Synoptic Scale Fronts from Atmospheric Data Grids. Weather. Clim. Dyn. Discuss. [Preprint] 2021 , 1–28. [ Google Scholar ] [ CrossRef ]
• Knighton, J.; Pleiss, G.; Carter, E.; Lyon, S.; Walter, M.T.; Steinschneider, S. Potential Predictability of Regional Precipitation and Discharge Extremes Using Synoptic-Scale Climate Information via Machine Learning: An Evaluation for the Eastern Continental United States. J. Hydrometeorol. 2019 , 20 , 883–900. [ Google Scholar ] [ CrossRef ]
• Hu, H.; Ayyub, B.M. Machine Learning for Projecting Extreme Precipitation Intensity for Short Durations in a Changing Climate. Geosciences 2019 , 9 , 209. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Elsanabary, M.H.; Gan, T.Y. Wavelet Analysis of Seasonal Rainfall Variability of the Upper Blue Nile Basin, Its Teleconnection to Global Sea Surface Temperature, and Its Forecasting by an Artificial Neural Network. Mon. Weather Rev. 2014 , 142 , 1771–1791. [ Google Scholar ] [ CrossRef ]
• Ghaderpour, E.; Vujadinovic, T.; Hassan, Q.K. Application of the Least-Squares Wavelet Software in Hydrology: Athabasca River Basin. J. Hydrol. Reg. Stud. 2021 , 36 , 100847. [ Google Scholar ] [ CrossRef ]
• Coulibaly, P.; Burn, D.H. Wavelet Analysis of Variability in Annual Canadian Streamflows. Water Resour. Res. 2004 , 40 . [ Google Scholar ] [ CrossRef ]
• Rolnick, D.; Donti, P.L.; Kaack, L.H.; Kochanski, K.; Lacoste, A.; Sankaran, K.; Ross, A.S.; Milojevic-Dupont, N.; Jaques, N.; Waldman-Brown, A.; et al. Tackling Climate Change with Machine Learning. arXiv 2019 , arXiv:1906.05433. [ Google Scholar ]
• Schneider, T.; Lan, S.; Stuart, A.; Teixeira, J. Earth system modeling 2.0: A blueprint for models that learn from observations and targeted high-resolution simulations. Geophys. Res. Lett. 2017 , 44 , 12396–12417. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Reichstein, M.; Camps-Valls, G.; Stevens, B.; Jung, M.; Denzler, J.; Carvalhais, N.; Prabhat. Deep learning and process understanding for data-driven Earth system science. Nature 2019 , 566 , 195–204. [ Google Scholar ] [ CrossRef ]
• Huntingford, C.; Jeffers, E.S.; Bonsall, M.B.; Christensen, H.M.; Lees, T.; Yang, H. Machine Learning and Artificial Intelligence to Aid Climate Change Research and Preparedness. Environ. Res. Lett. 2019 , 14 , 124007. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Dijkstra, H.A.; Petersik, P.; Hernández-García, E.; López, C. The Application of Machine Learning Techniques to Improve El Niño Prediction Skill. Front. Phys. 2019 , 7 , 153. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Mohammadiziazi, R.; Bilec, M.M. Application of Machine Learning for Predicting Building Energy Use at Different Temporal and Spatial Resolution under Climate Change in USA. Buildings 2020 , 10 , 139. [ Google Scholar ] [ CrossRef ]
• Piryonesi, S.M.; El-Diraby, T. Climate Change Impact on Infrastructure: A Machine Learning Solution for Predicting Pavement Condition Index. Constr. Build. Mater. 2021 , 306 , 124905. [ Google Scholar ] [ CrossRef ]
• Leng, G.; Hall, J.W. Predicting Spatial and Temporal Variability in Crop Yields: An Inter-Comparison of Machine Learning, Regression and Process-Based Models. Environ. Res. Lett. 2020 , 15 , 044027. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Feng, P.; Wang, B.; Liu, D.L.; Waters, C.; Yu, Q. Incorporating Machine Learning with Biophysical Model Can Improve the Evaluation of Climate Extremes Impacts on Wheat Yield in South-Eastern Australia. Agric. For. Meteorol. 2019 , 275 , 100–113. [ Google Scholar ] [ CrossRef ]
• Yu, T.; Zhang, Q.; Sun, R. Comparison of Machine Learning Methods to Up-Scale Gross Primary Production. Remote Sens. 2021 , 13 , 2448. [ Google Scholar ] [ CrossRef ]
• Duan, Z.; Yang, Y.; Zhou, S.; Gao, Z.; Zong, L.; Fan, S.; Yin, J. Estimating Gross Primary Productivity (GPP) over Rice–Wheat-Rotation Croplands by Using the Random Forest Model and Eddy Covariance Measurements: Upscaling and Comparison with the MODIS Product. Remote Sens. 2021 , 13 , 4229. [ Google Scholar ] [ CrossRef ]
• Zeng, J.; Matsunaga, T.; Tan, Z.-H.; Saigusa, N.; Shirai, T.; Tang, Y.; Peng, S.; Fukuda, Y. Global Terrestrial Carbon Fluxes of 1999–2019 Estimated by Upscaling Eddy Covariance Data with a Random Forest. Sci. Data 2020 , 7 , 313. [ Google Scholar ] [ CrossRef ]
• Zeng, Z.; Wang, Z.; Gui, K.; Yan, X.; Gao, M.; Luo, M.; Geng, H.; Liao, T.; Li, X.; An, J.; et al. Daily Global Solar Radiation in China Estimated From High-Density Meteorological Observations: A Random Forest Model Framework. Earth Space Sci. 2020 , 7 , e2019EA001058. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Yang, Y.; Zhang, M.; Li, Q.; Chen, B.; Gao, Z.; Ning, G.; Liu, C.; Li, Y.; Luo, M. Modulations of Surface Thermal Environment and Agricultural Activity on Intraseasonal Variations of Summer Diurnal Temperature Range in the Yangtze River Delta of China. Sci. Environ. 2020 , 736 , 139445. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Li, X.; Yang, Y.; Mi, J.; Bi, X.; Zhao, Y.; Huang, Z.; Liu, C.; Zong, L.; Li, W. Leveraging Machine Learning for Quantitative Precipitation Estimation from Fengyun-4 Geostationary Observations and Ground Meteorological Measurements. Atmos. Meas. Tech. 2021 , 14 , 7007–7023. [ Google Scholar ] [ CrossRef ]
• Ustrnul, Z. Zmienność Cyrkulacji Atmosfery na Półkuli Północnej w XX Wieku (Variability of the Atmospheric Circulation on the Northern Hemisphere in the 20th Century) ; Materiały Badawcze, seria Meteorologia; IMGW: Warszawa, Poland, 1997; Volume 27, p. 208, (In Polish, Eng. Summary). [ Google Scholar ]
• Tveito, O.E.; Ustrnul, Z. A Review of The Use of Large-Scale Atmospheric Circulation Classification in Spatial Climatology ; KLIMA; DNMI: Oslo, Norway, 3 October 2003; p. 27. [ Google Scholar ]
• Huth, R.; Beck, C.; Philipp, A.; Demuzere, M.; Ustrnul, Z.; Cahynová, M.; Kyselý, J.; Tveito, O.E. Classifications of atmospheric circulation patterns: Recent advances and applications. Ann. N. Y. Acad. Sci. 2008 , 1146 , 105–152. [ Google Scholar ] [ CrossRef ]
• Ustrnul, Z.; Wypych, A.; Czekierda, D. Composite circulation index of weather extremes (the example for Poland). Meteorol. Z. 2013 , 22 , 551–559. [ Google Scholar ] [ CrossRef ]
• Gordon, A.D. Classification ; Chapman & Hall: London, UK; New York, NY, USA; Washington, DC, USA, 1999. [ Google Scholar ]
• Everitt, B.S.; Landau, S.; Leese, M. Cluster Analysis ; Arnold: London, UK; Oxford University Press: New York, NY, USA, 2001. [ Google Scholar ]
• Bednorz, E.; Kolendowicz, L.; Bielec-Bąkowska, Z.; Bokwa, A.; Żelazny, M.; Kicińska, B.; Lewik, P.; Nowosad, M.; Ustrnul, Z. Regionalizacje, typologie i wydzielanie sezonów klimatycznych z zastosowaniem analizy skupień (Regionalisations, typologies and climatic seasons with application of cluster analysis). Prz. Geofiz. 2003 , 1–2 , 11–32, (In Polish, Eng. Summary). [ Google Scholar ]
• Sheridan, S.C.; Lee, C.C. The self-organizing map in synoptic climatological research. Prog. Phys. Geogr. Earth Environ. 2011 , 35 , 109–119. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Lee, C.C.; Sheridan, S.C. A six-step approach to developing future synoptic classifications based on GCM output. Int. J. Climatol. 2012 , 32 , 1792–1802. [ Google Scholar ] [ CrossRef ]
• Davis, R.E.; Kalkstein, L.S. Development of an automated spatial synoptic classification. Int. J. Climatol. 1990 , 10 , 769–794. [ Google Scholar ] [ CrossRef ]
• Wilks, S.D. Statistical Methods in the Atmospheric Sciences, an Introduction ; Academic Press: San Diego, CA, USA, 1995; p. 464. [ Google Scholar ]
• Oliver, J.E. The history, status, and future of climatic classifications. Phys. Geogr. 1991 , 12 , 231–251. [ Google Scholar ] [ CrossRef ]
• Yarnal, B. Synoptic Climatology in Environmental Analysis: A Primer ; Belhaven Press: London, UK, 1993; pp. xv + 195. [ Google Scholar ]
• Barry, R.G.; Carleton, A.M. Synoptic and Dynamic Climatology ; Routledge: London, UK, 2001; p. 664. [ Google Scholar ] [ CrossRef ]
• Yarnal, B.; Comrie, A.C.; Frakes, B.; Brown, D.P. Developments and prospects in synoptic climatology. Int. J. Climatol. 2001 , 21 , 1923–1950. [ Google Scholar ] [ CrossRef ]
• Palarz, A.; Luterbacher, J.; Ustrnul, Z.; Xoplaki, E.; Celiński-Myslaw, D. Representation of low-tropospheric temperature inversions in ECMWF reanalyses over Europe. Environ. Res. Lett. 2020 , 15 , 074043. [ Google Scholar ] [ CrossRef ]
• Hertz, J.; Krogh, A.; Palmer, R.G. Introduction to the Theory of Neural Computation ; Addison-Wesley Publ. Company: Reading, MA, USA, 1991. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Pasini, A.; Pelino, V.; Potesta, S. A neural network model for visibility nowcasting from surface observations: Results and sensitivity to physical input variables. J. Geophys. Res. Atmos. 2001 , 106 , 14951–14960. [ Google Scholar ] [ CrossRef ]
• Domańska, D.; Wojtylak, M. Application of fuzzy time series models for forecasting pollution concentrations. Expert. Syst. Appl. 2012 , 39 , 7673–7679. [ Google Scholar ] [ CrossRef ]
• Czernecki, B.; Marosz, M.; Jędruszkiewicz, J. Assessment of Machine Learning Algorithms in Short-Term Forecasting of PM10 and PM2.5 Concentrations in Selected Polish Agglomerations. Aerosol Air Qual. Res. 2021 , 21 , 200586. [ Google Scholar ] [ CrossRef ]
• Zalewska, T.; Biernacik, D.; Marosz, M. Correlations between 7Be, 210Pb, Dust and PM10 Concentrations in Relation to Meteorological Conditions in Northern Poland in 1998–2018. J. Environ. Radioact. 2021 , 228 , 106526. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Dixon, P.G.; Allen, M.; Gosling, S.N.; Hondula, D.M.; Vijendra, I.; Lucas, R.; Vanos, J. Perspectives on the Synoptic Climate Classification and its Role in Interdisciplinary Research. Geogr. Compass 2016 , 10 , 147–164. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Bauer, P.; Dueben, P.D.; Hoefler, T.; Quintino, T.; Schulthess, T.C.; Wedi, N.P. The Digital Revolution of Earth-System Science. Nat. Comput. Sci. 2021 , 1 , 104–113. [ Google Scholar ] [ CrossRef ]
• Dewitte, S.; Cornelis, J.P.; Müller, R.; Munteanu, A. Artificial Intelligence Revolutionises Weather Forecast, Climate Monitoring and Decadal Prediction. Remote Sens. 2021 , 13 , 3209. [ Google Scholar ] [ CrossRef ]
• Mock, C.J.; Shinker, J.J. Paleoclimate—Modern Analog Approaches in Paleoclimatology. In Paleoclimate, Encyclopedia of Quaternary Science , 2nd ed.; Elias, S.A., Mock, C.J., Eds.; Elsevier: Amsterdam, The Netherlands, 2013; pp. 102–112. [ Google Scholar ]

Click here to enlarge figure

Share and Cite

Bochenek, B.; Ustrnul, Z. Machine Learning in Weather Prediction and Climate Analyses—Applications and Perspectives. Atmosphere 2022 , 13 , 180. https://doi.org/10.3390/atmos13020180

Bochenek B, Ustrnul Z. Machine Learning in Weather Prediction and Climate Analyses—Applications and Perspectives. Atmosphere . 2022; 13(2):180. https://doi.org/10.3390/atmos13020180

Bochenek, Bogdan, and Zbigniew Ustrnul. 2022. "Machine Learning in Weather Prediction and Climate Analyses—Applications and Perspectives" Atmosphere 13, no. 2: 180. https://doi.org/10.3390/atmos13020180

Article Metrics

Article access statistics, supplementary material.

ZIP-Document (ZIP, 346 KiB)

Further Information

Mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

Advertisement

Big Data Analytics in Weather Forecasting: A Systematic Review

• Review article
• Published: 28 June 2021
• Volume 29 , pages 1247–1275, ( 2022 )

Cite this article

• Marzieh Fathi 1 ,
• Mostafa Haghi Kashani 2 ,
• Seyed Mahdi Jameii   ORCID: orcid.org/0000-0002-9407-665X 2 &
• Ebrahim Mahdipour 1  

8224 Accesses

73 Citations

Explore all metrics

A Correction to this article was published on 20 July 2021

This article has been updated

Weather forecasting, as an important and indispensable procedure in people’s daily lives, evaluates the alteration happening in the current condition of the atmosphere. Big data analytics is the process of analyzing big data to extract the concealed patterns and applicable information that can yield better results. Nowadays, several parts of society are interested in big data, and the meteorological institute is not excluded. Therefore, big data analytics will give better results in weather forecasting and will help forecasters to forecast weather more accurately. In order to achieve this goal and to recommend favorable solutions, several big data techniques and technologies have been suggested to manage and analyze the huge volume of weather data from different resources. By employing big data analytics in weather forecasting, the challenges related to traditional data management techniques and technology can be solved. This paper tenders a systematic literature review method for big data analytic approaches in weather forecasting (published between 2014 and August 2020). A feasible taxonomy of the current reviewed papers is proposed as technique-based, technology-based, and hybrid approaches. Moreover, this paper presents a comparison of the aforementioned categories regarding accuracy, scalability, execution time, and other Quality of Service factors. The types of algorithms, measurement environments, modeling tools, and the advantages and disadvantages per paper are extracted. In addition, open issues and future trends are debated.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Similar content being viewed by others

Big Data Analytics Using Data Mining Techniques: A Survey

Time series big data: a survey on data stream frameworks, analysis and algorithms

Big data and predictive analytics: A systematic review of applications

Explore related subjects.

• Artificial Intelligence

Change history

20 july 2021.

A Correction to this paper has been published: https://doi.org/10.1007/s11831-021-09630-6

Xiao Z, Liu B, Liu H, Zhang D (2012) Progress in climate prediction and weather forecast operations in China. Adv Atmos Sci 29(5):943–957

Google Scholar  

Bengtsson L (1980) The weather forecast. Pure Appl Geophys 119(3):515–537

Kan L, Yu-Shu L (2005) A rough set based fuzzy neural network algorithm for weather prediction. In: 2005 International conference on machine learning and cybernetics, vol 3. pp 1888–1892

Kan L, Yu-Shu L (2002) Fuzzy case-based reasoning: weather prediction. In: Proceedings of the international conference on machine learning and cybernetics, vol 1. pp 107–110

Weiguo X (2010) The weather prediction method based on artificial immune system. In: 2010 International forum on information technology and applications, vol 2. pp 386–389

Haupt SE, Cowie J, Linden S, McCandless T, Kosovic B, Alessandrini S (2018) Machine learning for applied weather prediction. In: 2018 IEEE 14th international conference on e-science (e-Science). pp 276–277

Chung CYC, Kumar VR (1993) Knowledge acquisition using a neural network for a weather forecasting knowledge-based system. Neural Comput Appl 1(3):215–223

Pandey AK, Agrawal CP, Agrawal M (2017) A hadoop based weather prediction model for classification of weather data. In: 2017 Second international conference on electrical, computer and communication technologies (ICECCT). pp 1–5

Tsai C-W, Lai C-F, Chao H-C, Vasilakos AV (2015) Big data analytics: a survey. J Big Data 2(1):21

Rodríguez-Mazahua L, Rodríguez-Enríquez C-A, Sánchez-Cervantes JL, Cervantes J, García-Alcaraz JL, Alor-Hernández G (2016) A general perspective of big data: applications, tools, challenges and trends. J Supercomput 72(8):3073–3113

Talia D (2013) Clouds for scalable big data analytics. Computer 46(5):98–101

Selvaraj P, Marudappa P (2018) A survey of predictive analytics using big data with data mining. Int J Bioinf Res Appl 14:269

Sharma S, Mangat V (2015) Technology and trends to handle big data: survey. In: 2015 Fifth international conference on advanced computing and communication technologies. pp 266–271

Jain H, Jain R (2017) Big data in weather forecasting: applications and challenges. In: 2017 International conference on big data analytics and computational intelligence (ICBDAC). pp 138–142

Reddy PC, Babu AS (2017) Survey on weather prediction using big data analystics. In: 2017 Second international conference on electrical, computer and communication technologies (ICECCT). pp 1–6

Bendre MR, Thool RC, Thool VR (2015) Big data in precision agriculture: weather forecasting for future farming. In: 2015 1st international conference on next generation computing technologies (NGCT). pp 744–750

Mittal S, Sangwan OP (2019) Big data analytics using data mining techniques: a survey. In: Advanced informatics for computing research, Singapore. Springer Singapore, pp 264–273

Leu J-S, Su K-W, Chen C-T (2014) Ambient mesoscale weather forecasting system featuring mobile augmented reality. Multimed Tools Appl 72(2):1585–1609

Corne D, Dissanayake M, Peacock A, Galloway S, Owens E (2014) Accurate localized short term weather prediction for renewables planning. In: 2014 IEEE symposium on computational intelligence applications in smart grid (CIASG). pp 1–8

Roudier P et al (2014) The role of climate forecasts in smallholder agriculture: lessons from participatory research in two communities in Senegal. Clim Risk Mana 2:42–55

Li J, Xu L, Tang L, Wang S, Li L (2018) Big data in tourism research: a literature review. Tour Manag 68:301–323

Scott D, Lemieux C (2010) Weather and climate information for tourism. Procedia Environ Sci 1:146–183

Hazyuk I, Ghiaus C, Penhouet D (2012) Optimal temperature control of intermittently heated buildings using Model Predictive Control: Part I—Building modeling. Build Environ 51:379–387

Enríquez R, Jiménez MJ, Heras MdR (2016) Solar forecasting requirements for buildings MPC. Energy Procedia 91:1024–1032

Smith DA, Sherry L (2008) Decision support tool for predicting aircraft arrival rates from weather forecasts. In: 2008 Integrated communications, navigation and surveillance conference. pp 1–12

Zhang B, Tang L, Roemer M (2018) Probabilistic planning and risk evaluation based on ensemble weather forecasting. IEEE Trans Autom Sci Eng 15(2):556–566

Braman LM, van Aalst MK, Mason SJ, Suarez P, Ait-Chellouche Y, Tall A (2013) Climate forecasts in disaster management: red cross flood operations in West Africa, 2008. Disasters 37(1):144–164

Akhand MH (2003) Disaster management and cyclone warning system in Bangladesh. In: Zschau J, Küppers A (eds) Early warning systems for natural disaster reduction. Springer, Berlin, pp 49–64

Chen C, Duan S, Cai T, Liu B (2011) Online 24-h solar power forecasting based on weather type classification using artificial neural network. Sol Energy 85(11):2856–2870

Shi J, Lee W, Liu Y, Yang Y, Wang P (2012) Forecasting power output of photovoltaic systems based on weather classification and support vector machines. IEEE Trans Ind Appl 48(3):1064–1069

Lazos D, Sproul AB, Kay M (2014) Optimisation of energy management in commercial buildings with weather forecasting inputs: a review. Renew Sustain Energy Rev 39:587–603

Casas DM, González JÁT, Rodríguez JEA, Pet JV (2009) Using data-mining for short-term rainfall forecasting. In: Distributed computing, artificial intelligence, bioinformatics, soft computing, and ambient assisted living. Springer, Berlin, pp 487–490

Katal A, Wazid M, Goudar RH (2013) Big data: issues, challenges, tools and good practices. In: 2013 Sixth international conference on contemporary computing (IC3). pp 404–409

Elgendy N, Elragal A (2014) Big data analytics: a literature review paper. In: Advances in data mining. Applications and theoretical aspects. Springer, Cham, pp 214–227

Shadroo S, Rahmani A (2018) Systematic survey of big data and data mining in internet of things. Comput Netw 139:19–47

Bazzaz Abkenar S, Mahdipour E, Jameii SM, Haghi Kashani M (2021) A hybrid classification method for Twitter spam detection based on differential evolution and random forest. Concurr Comput Pract Exp. https://doi.org/10.1002/cpe.6381

Article   Google Scholar  

Pathak AR, Pandey M, Rautaray S (2018) Construing the big data based on taxonomy, analytics and approaches. Iran J Comput Sci 1(4):237–259

Bazzaz Abkenar S, Haghi Kashani M, Mahdipour E, Jameii SM (2021) Big data analytics meets social media: a systematic review of techniques, open issues, and future directions. Telemat Inform 57:101517

Khezr SN, Navimipour NJ (2017) MapReduce and its applications, challenges, and architecture: a comprehensive review and directions for future research. J Grid Comput 15(3):295–321

Amer A-B, Amr M, Salah H (2016) A survey on MapReduce implementations. Int J Cloud Appl Comput IJCAC 6(1):59–87

Senger H et al (2016) BSP cost and scalability analysis for MapReduce operations. Concurr Comput Pract Exp 28(8):2503–2527

Lee D, Kim JW, Maeng S (2014) Large-scale incremental processing with MapReduce. Future Gener Comput Syst 36:66–79

Idris M et al (2015) Context-aware scheduling in MapReduce: a compact review. Concurr Comput Pract Exp 27(17):5332–5349

Karimi Y, Haghi Kashani M, Akbari M, Mahdipour E (2021) Leveraging big data in smart cities: a systematic review. J Concurr Comput Pract Exp. https://doi.org/10.1002/cpe.6379

Bakratsas M, Basaras P, Katsaros D, Tassiulas L (2018) Hadoop MapReduce performance on SSDs for analyzing social networks. Big Data Res 11:1–10

Shabestari F, Rahmani AM, Navimipour NJ, Jabbehdari S (2019) A taxonomy of software-based and hardware-based approaches for energy efficiency management in the Hadoop. J Netw Comput Appl 126:162–177

Patwardhan A, Verma AK, Kumar U (2016) A survey on predictive maintenance through big data. In: Current trends in reliability, availability, maintainability and safety. Springer, Cham, pp 437–445

Yang W, Liu X, Zhang L, Yang LT (2013) Big data real-time processing based on storm. In: 2013 12th IEEE international conference on trust, security and privacy in computing and communications. pp 1784–1787

Philip-Chen CL, Zhang C-Y (2014) Data-intensive applications challenges techniques and technologies: a survey on big data. Inf Sci 275:314–347

Lee J, Hong S, Lee J-H (2014) An efficient prediction for heavy rain from big weather data using genetic algorithm. In: Presented at the proceedings of the 8th international conference on ubiquitous information management and communication, Siem Reap, Cambodia

Sahasrabuddhe DV, Jamsandekar P (2015) Data structure for representation of big data of weather forecasting: a review. Int J Comput Sci Trends Technol IJCST 3(6):48–56

Priya SB A survey on weather forecasting to predict rainfall using big data analytics

Hassani H, Silva ES (2015) Forecasting with big data: a review. Ann Data Sci 2(1):5–19

Rao N (2017) Big data and climate smart agriculture-review of current status and implications for agricultural research and innovation in India. In: Proceedings Indian National Science Academy, Forthcoming

de Freitas Viscondi G, Alves-Souza SN (2019) A systematic literature review on big data for solar photovoltaic electricity generation forecasting. Sustain Energy Technol Assess 31:54–63

Vannitsem S et al (2021) Statistical postprocessing for weather forecasts: review, challenges, and avenues in a big data world. Bull Am Meteorol Soc 102(3):E681–E699

Cook DJ, Greengold NL, Ellrodt AG, Weingarten SR (1997) The relation between systematic reviews and practice guidelines. Ann Internal Med 127(3):210–216

Haghi Kashani M, Rahmani AM, Jafari Navimipour N (2020) Quality of service-aware approaches in fog computing. Int J Commun Syst 33(8):e4340

Rahimi M, Songhorabadi M, Haghi Kashani M (2020) Fog-based smart homes: a systematic review. J Netw Comput Appl 153:102531

Bazzaz Abkenar S, Haghi Kashani M, Akbari M, Mahdipour E (2020) Twitter spam detection: a systematic review. arXiv preprint arXiv: 2011.14754 .

Songhorabadi M, Rahimi M, Farid AMM, Kashani MH (2020) Fog computing approaches in smart cities: a state-of-the-art review. arXiv preprint arXiv: 2011.14732

Kashani MH, Ahmadzadeh A, Mahdipour E (2020) Load balancing mechanisms in fog computing: a systematic review. arXiv preprint arXiv: 2011.14706

Brereton P, Kitchenham BA, Budgen D, Turner M, Khalil M (2007) Lessons from applying the systematic literature review process within the software engineering domain. J Syst Softw 80(4):571–583

Sheikh Sofla M, Haghi Kashani M, Mahdipour E, Faghih Mirzaee R (2021) Towards effective offloading mechanisms in fog computing: a systematic survey. Multimed Tools Appl

Haghi Kashani M, Madanipour M, Nikravan M, Asghari P, Mahdipour E (2021) A systematic review of IoT in healthcare: applications, techniques, and trends. J Netw Comput Appl

Cheng Y, Zheng Z, Wang J, Yang L, Wan S (2019) Attribute reduction based on genetic algorithm for the coevolution of meteorological data in the industrial internet of things. Wirel Commun Mob Comput 2019:8

Cramer S, Kampouridis M, Freitas A (2016) A genetic decomposition algorithm for predicting rainfall within financial weather derivatives. In: Presented at the proceedings of the genetic and evolutionary computation conference 2016, Denver, Colorado, USA

Pooja SB, Siva-Balan RV, Anisha M, Muthukumaran MS, Jothikumar R (2020) Techniques Tanimoto correlated feature selection system and hybridization of clustering and boosting ensemble classification of remote sensed big data for weather forecasting. Comput Commun 151:266–274

Kvinge H, Farnell E, Kirby M, Peterson C (2018) Monitoring the shape of weather, soundscapes, and dynamical systems: a new statistic for dimension-driven data analysis on large datasets. In: 2018 IEEE international conference on big data (big data). pp 1045–1051

Buszta A, Mazurkiewicz J (2015) Climate changes prediction system based on weather big data visualisation. In: Theory and engineering of complex systems and dependability. Springer, Cham, pp 75–86

Rasel RI, Sultana N, Meesad P (2018) An application of data mining and machine learning for weather forecasting. In: Recent advances in information and communication technology 2017. Springer, Cham, pp 169–178

Mahmood MR, Patra RK, Raja R, Sinha GR (2019) A novel approach for weather prediction using forecasting analysis and data mining techniques. In: Innovations in electronics and communication engineering. Springer, Singapore, pp 479–489

Azimi R, Ghofrani M, Ghayekhloo M (2016) A hybrid wind power forecasting model based on data mining and wavelets analysis. Energy Convers Manag 127:208–225

Doreswamy IG, Manjunatha BR (2018) Multi-label classification of big NCDC weather data using deep learning model. In: Soft computing systems. Springer, Singapore, pp 232–241

Venkatachalapathy K, Kamaleshwar T, Sundaranarayana D, Prakash VO (2016) An effective framework with N-client transfer dataset for weather prediction using data mining techniques. In: Presented at the proceedings of the international conference on informatics and analytics, Pondicherry, India

Choi C, Kim J, Kim J, Kim D, Bae Y, Kim HS (2018) Development of heavy rain damage prediction model using machine learning based on big data. Adv Meteorol 2018:11

Hubig N, Fengler P, Züfle A, Yang R, Günnemann S (2017) Detection and prediction of natural hazards using large-scale environmental data. In: Advances in spatial and temporal databases. Springer, Cham, pp 300–316

Yonekura K, Hattori H, Suzuki T (2018) Short-term local weather forecast using dense weather station by deep neural network. In: 2018 IEEE international conference on big data (big data). pp 1683–1690

Xu Q et al (2015) A short-term wind power forecasting approach with adjustment of numerical weather prediction input by data mining. IEEE Trans Sustain Energy 6(4):1283–1291

Jiang P, Dong Q (2015) A new hybrid model based on an intelligent optimization algorithm and a data denoising method to make wind speed predication. Math Probl Eng 2015:16

More PD, Nandgave S, Kadam M (2020) Weather data analytics using hadoop with map-reduce. In: ICCCE 2019. Springer, Singapore, pp 189–196

Wu H (2017) Big data management the mass weather logs. In: Smart computing and communication. Springer, Cham, pp 122–132

Ismail KA, Majid MA, Zain JM, Bakar NAA (2016) Big data prediction framework for weather temperature based on MapReduce algorithm. In: 2016 IEEE conference on open systems (ICOS). pp 13–17

Abdullahi AU, Ahmad R, Zakaria NM (2016) Big data: performance profiling of meteorological and oceanographic data on hive. In: 2016 3rd international conference on computer and information sciences (ICCOINS). pp 203–208

Oury DTM, Singh A (2018) Data analysis of weather data using hadoop technology. In: Smart computing and informatics. Springer, Singapore, pp 723–730

Manogaran G, Lopez D, Chilamkurti N (2018) In-mapper combiner based MapReduce algorithm for processing of big climate data. Future Gener Comput Syst 86:433–445

Jayanthi D, Sumathi G (2017) Weather data analysis using spark—an in-memory computing framework. In: 2017 Innovations in power and advanced computing technologies (i-PACT). pp 1–5

Palamuttam R et al (2015) SciSpark: Applying in-memory distributed computing to weather event detection and tracking. In: 2015 IEEE International conference on big data (big data). pp 2020–2026

Hassaan M, Elghandour I (2016) A real-time big data analysis framework on a CPU/GPU heterogeneous cluster: a meteorological application case study. In: 2016 IEEE/ACM 3rd international conference on big data computing applications and technologies (BDCAT). pp 168–177

Manogaran G, Lopez D (2018) Spatial cumulative sum algorithm with big data analytics for climate change detection. Comput Electr Eng 65:207–221

Madan S, Kumar P, Rawat S, Choudhury T (2018) Analysis of weather prediction using machine learning & big data. In: 2018 International conference on advances in computing and communication engineering (ICACCE). pp 259–264

Dhoot R, Agrawal S, Kumar MS (2019) Implementation and analysis of arima model and kalman filter for weather forcasting in spark computing environment. In: 2019 3rd international conference on computing and communications technologies (ICCCT). pp 105–112

Dhamodharavadhani S, Rathipriya R (2019) Region-wise rainfall prediction using mapreduce-based exponential smoothing techniques. In: Advances in big data and cloud computing. Springer, Singapore, pp 229–239

Namitha K, Jayapriya A, Kumar GS (2015) Rainfall prediction using artificial neural network on map-reduce framework. In: Presented at the proceedings of the third international symposium on women in computing and informatics, Kochi, India

Liu L, Lv J, Ma Z, Wan J, Jingjing M (2015) Toward the association rules of meteorological data mining based on cloud computing. In: Proceedings of the second international conference on mechatronics and automatic control. Springer, Cham, pp 1051–1059

Sahoo S (2017) A parallel forecasting approach using incremental K-means clustering technique. In: Computational intelligence in data mining. Springer, Singapore, pp 165–172

Fang W, Sheng VS, Wen X, Pan W (2014) Meteorological data analysis using MapReduce. Sci World J 2014:10

Hamzei M, Navimipour NJ (2018) Toward efficient service composition techniques in the internet of things. IEEE Internet Things J 5(5):3774–3787

Kumar V, Kumar D (2020) A systematic review on firefly algorithm: past, present, and future. Arch Comput Methods Eng 28(4):3269–3291

MathSciNet   Google Scholar  

Nikravan M, Kashani MH (2007) Parallel min–max ant colony system (MMAS) for dynamic process scheduling in distributed operating systems considering load balancing. In: Proceedings of the 21st ECMS international conference on high performance computing & simulation (HPCS), Prague, Czech Republic

Kashani MH, Sarvizadeh R (2011) A novel method for task scheduling in distributed systems using max–min ant colony optimization. In: 2011 3rd international conference on advanced computer control (ICACC). IEEE, pp 422–426

Kashani MH, Zarrabi H, Javadzadeh G (2017) A new metaheuristic approach to task assignment problem in distributed systems. In: 2017 IEEE 4th international conference on knowledge-based engineering and innovation (KBEI). IEEE, pp 0673–0677

Kashani MH, Sarvizadeh R, Jameii M (2012) A new distributed systems scheduling algorithm: a swarm intelligence approach. In: Fourth international conference on machine vision (ICMV 2011): computer vision and image analysis; pattern recognition and basic technologies. International Society for Optics and Photonics

Kashani MH, Jahanshahi M (2009) A new method based on memetic algorithm for task scheduling in distributed systems. Int J Simul Syst Sci Technol 10

Niu B, Wang H (2012) Bacterial colony optimization. Discrete Dyn Nat Soc 2012:28

MathSciNet   MATH   Google Scholar  

Kashani MH, Jahanshahi M (2009) Using simulated annealing for task scheduling in distributed systems. In: 2009 International conference on computational intelligence, modelling and simulation. pp 265–269

Dasgupta D, Ji Z, Gonzalez F (2003) Artificial immune system (AIS) research in the last five years. In: The 2003 congress on evolutionary computation, 2003. CEC '03., vol 1. pp 123–130

Jameii SM, Kashani MH, Karimi R (2015) LASPEA: Learning automata-based strength pareto evolutionary algorithm for multi-objective optimization. Int J Comput Sci Telecommun 6(9):14–19

Yang X-S (2010) A new metaheuristic bat-inspired algorithm. In: Nature inspired cooperative strategies for optimization (NICSO 2010). Springer, pp 65–74

Yang X-S. Bat algorithm for multi-objective optimisation. arXiv e-prints, Accessed 01 Mar 2012. arXiv:1203.6571Y

Krishnanand KN, Ghose D (2009) Glowworm swarm optimization for simultaneous capture of multiple local optima of multimodal functions. Swarm Intell 3(2):87–124

Sarvizadeh R, Kashani MH, Zakeri FS, Jameii SM (2012) A novel bee colony approach to distributed systems scheduling. Int J Comput Appl 42(10):1–6

Saneja B, Rani R (2018) A hybrid approach for outlier detection in weather sensor data. In: 2018 IEEE 8th international advance computing conference (IACC). pp 321–326

Al-Madi N, Aljarah I, Ludwig S (2014) Parallel Glowworm Swarm Optimization Clustering Algorithm based on MapReduce

El-Alfy E-SM, Alshammari MA (2016) Towards scalable rough set based attribute subset selection for intrusion detection using parallel genetic algorithm in MapReduce. Simul Model Pract Theory 64(13):18–29

Download references

Author information

Authors and affiliations.

Department of Computer Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran

Marzieh Fathi & Ebrahim Mahdipour

Department of Computer Engineering, Shahr-e-Qods Branch, Islamic Azad University, Tehran, Iran

Mostafa Haghi Kashani & Seyed Mahdi Jameii

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Seyed Mahdi Jameii .

Ethics declarations

Conflict of interest.

On behalf of all authors, the corresponding author states that there is no conflict of interest.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Fathi, M., Haghi Kashani, M., Jameii, S.M. et al. Big Data Analytics in Weather Forecasting: A Systematic Review. Arch Computat Methods Eng 29 , 1247–1275 (2022). https://doi.org/10.1007/s11831-021-09616-4

Download citation

Received : 27 December 2020

Accepted : 12 June 2021

Published : 28 June 2021

Issue Date : March 2022

DOI : https://doi.org/10.1007/s11831-021-09616-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Find a journal
• Publish with us
• Track your research

Weather Research & Forecasting Model (WRF)

A state of the art mesoscale numerical weather prediction system designed for both atmospheric research and operational forecasting applications

Current Version 4.5

The Weather Research and Forecasting (WRF) Model is a state of the art mesoscale numerical weather prediction system designed for both atmospheric research and operational forecasting applications. It features two dynamical cores, a data assimilation system, and a software architecture supporting parallel computation and system extensibility. The model serves a wide range of meteorological applications across scales from tens of meters to thousands of kilometers. The effort to develop WRF began in the latter 1990's and was a collaborative partnership of the National Center for Atmospheric Research (NCAR), the National Oceanic and Atmospheric Administration (represented by the National Centers for Environmental Prediction (NCEP) and the Earth System Research Laboratory), the U.S. Air Force, the Naval Research Laboratory, the University of Oklahoma, and the Federal Aviation Administration (FAA).

For researchers, WRF can produce simulations based on actual atmospheric conditions (i.e., from observations and analyses) or idealized conditions. WRF offers operational forecasting a flexible and computationally-efficient platform, while reflecting recent advances in physics, numerics, and data assimilation contributed by developers from the expansive research community. WRF is currently in operational use at NCEP and other national meteorological centers as well as in real-time forecasting configurations at laboratories, universities, and companies.

WRF has a large worldwide community of registered users (a cumulative total of over 57,800 in over 160 countries as of 2021), and NCAR provides regular workshops and tutorials on it.

This site provides general background information on the WRF Model and its organization and offers links to information on user support, code contributions, and system administration. For detailed information on model use, updates and events, support, code downloads, and documentation, please visit the WRF-ARW users' page (see above).

For any questions regarding the WRF model, please post to the WRF & MPAS-A Support Forum .

PM Modi launches 3 indigenous Param Rudra supercomputers — know all about it

These supercomputers, built at a cost of ₹130 crore, have been deployed in pune, delhi and kolkata to facilitate pioneering scientific research, modi said, addressing scientists and others via video conferencing. pm modi also inaugurated the ₹850-crore high-performance computing system tailored for weather and climate research..

Prime Minister Narendra Modi on Thursday, September 26, launched three PARAM Rudra supercomputers, developed indigenously under the National Supercomputing Mission. The initiative marks a significant step in the country’s weather and climate research capabilities. These supercomputers, built at a cost of ₹130 crore, have been deployed in Pune, Delhi and Kolkata to facilitate pioneering scientific research, Modi said, addressing scientists and others via video conferencing. Modi also inaugurated the ₹850-crore high-performance computing system tailored for weather and climate research, and described the day as one of a "very big achievement" in the science and technology sector. "With Param Rudra Supercomputers and HPC system, India takes significant step towards self-reliance in computing and driving innovation in science and tech," Modi said at the launch. The launch, conducted at 5.30 pm today, also highlighted the importance of science and technology in shaping India’s future, with the prime minister urging the younger generation to take part in this major technological milestone. "A boost to tech related innovations! At around 5:30 PM this evening, via video conferencing, I will be inaugurating 3 Param Rudra Super Computing Systems and a High Performance Computing System for weather and climate. Will specially urge my young friends to join..." he said in an X post ahead of the launch. Speaking at the launch PM Modi said, “There is no sector which does not rely on technology and computing capability," adding, "In this revolution, our share should not be in bits and bytes but in terabytes and petabytes. Therefore, this achievement proves that we are moving in the right direction at the right pace." The significance of science is not only in invention and development, but also in fulfilling the aspirations of the last person, Modi said. What is Param Rudra and how will it revolutionise research? Developed by C-DAC Pune, the Param Rudra supercomputing system boasts an impressive 1-petaflop processing capability and features graphics processing units (GPUs) that enhance its ability to process vast amounts of data. This powerful system is set to revolutionise research in fields like radio astronomy and meteorology, making significant advancements in studying the interstellar medium and celestial phenomena. The system’s enhanced sensitivity to low and mid-radio frequencies will enable researchers to conduct more accurate and detailed studies, ultimately leading to breakthroughs in understanding cosmic phenomena. Param Rudra's capabilities will support the Giant Metre Radio Telescope (GMRT) in Pune, one of the world’s largest telescopes, improving its capacity to analyse Fast Radio Bursts (FRBs) and other astronomical phenomena. By processing more complex datasets with greater efficiency, the system will contribute to expanding knowledge in cosmology, black holes, and the origins of the universe. Located at two key sites, the Indian Institute of Tropical Meteorology (IITM) in Pune and the National Center for Medium Range Weather Forecast (NCMRWF) in Noida, this system has extraordinary computing power. Arka and Arunika — the two new high-power computing systems The new high-power computing systems — named Arka and Arunika — are expected to significantly enhance the accuracy and lead time of predictions related to tropical cyclones, heavy precipitation, thunderstorms, hailstorms, heat waves, droughts, and other critical weather phenomena. Arunika and Arka will replace the existing high-power computing systems, Pratyush and Mihir, located at NCMRWF, Noida and IITM, Pune, respectively. With a combined processing power of 21.3 petaflops, they will greatly improve India’s capacity for weather modelling and climate forecasting. The supercomputing systems feature advanced liquid cooling technology and robust data storage, ensuring optimal performance and long-term efficiency. These systems, based on Eviden's BullSequana XH2000, will have a combined power capacity of up to 21.3 petaflops and will rely on Eviden's patented direct liquid cooling technology to cool the system using warm water. The new systems at IITM Pune will provide 10.7 petaflops of power dedicated to atmospheric and climate research, while the NCMRWF facility in Noida will host a system with 7.49 petaflops of computing power to enhance weather and climate modeling efforts. This platform will bring together 2,115 CPU nodes with AMD Milan 7643 processors, 18 GPU nodes using NVIDIA A100 Tensor Core GPUs, the NVIDIA Quantum InfiniBand networking platform, and high-performance memory provided by Micron. The NCMRWF supercomputer will also have 2PB all flash and 20PB disk-based DDN EXAScaler ES400NVX2 parallel filesystem storage. The new systems at IITM Pune will integrate 3,021 CPU nodes using AMD Milan 7643 processors and 26 GPU nodes through NVIDIA A100 Tensor Core GPUs. Both systems will also foster global collaboration, with India set to become a key player in international meteorological research, attracting partnerships from scientists and research institutions worldwide. Also Read: Watch | PM Modi meets Chess Olympiad gold medallists at official residence A major leap in India's computational capabilities This ambitious project underscores India’s commitment to becoming self-reliant in supercomputing technology and improving its ability to forecast weather and climate conditions. With the combination of Param Rudra, Arka, and Arunika, India is poised to make significant contributions to scientific advancements in fields ranging from astronomy to weather prediction, and further cement its position as a leader in global research collaborations. The addition of the supercomputing system represents a significant upgrade for the Giant Metre Radio Telescope (GMRT), enhancing its computational power to process vast amounts of data with greater efficiency and precision. This upgrade is crucial for improving the telescope's sensitivity to low-frequency signals from space, enabling deeper insights into distant celestial bodies and cosmic phenomena. With the advanced capabilities of the Param Rudra system, astronomers can conduct more complex analyses of data, thereby accelerating breakthroughs in understanding the universe's origins, black holes, and the interstellar medium. Moreover, since its inception in 2002, the GMRT has been a vital part of the global astronomy community, and these upgraded capabilities will continue to attract collaborations from scientists worldwide, fostering international research partnerships in the field of radio astronomy. In addition, the enhanced computational capabilities provided by these systems will not only accelerate scientific breakthroughs but also have far-reaching implications for various sectors, including agriculture, disaster management, and public safety, making this a pivotal moment in India's technological progress.