Meet FourCastNet: a global data-driven weather forecasting model that revolutionizes weather predictions with a fast and accurate deep learning approach

In the 1920s, numerical weather prediction (NWP) emerged. They are ubiquitous and help with economic planning in major industries such as transportation, logistics, agriculture, and energy production. Many lives have been saved thanks to accurate weather forecasts that warned in advance of serious disasters. In recent decades, the quality of weather forecasts has improved. Lewis Fry Richardson used a slide rule and a table of logarithms to calculate the first dynamically modeled numerical weather prediction for a single location in 1922. It took him six weeks to produce a 6-hour forecast of the atmosphere. Early electronic computers significantly increased forecasting speed in the 1950s, allowing operational forecasts to be calculated quickly enough to be useful for future predictions.

In addition to improving computational power, improvements in weather prediction have been achieved thanks to better parameterization of small-scale phenomena through a deeper understanding of their physics and better atmospheric observations. By assimilating data, the latter has led to better model initializations. Because they have much cheaper processing costs than state-of-the-art NWP models, data-driven deep learning (DL) models are becoming increasingly popular for weather forecasting. Building data-driven models to predict large-scale circulation of the atmosphere has been the subject of several investigations. These models have been trained using outputs from climate models, general circulation models (GCMs), reanalysis products, or a combination of outputs from climate models and reanalysis products.

By eliminating biases prevalent in NWP models and enabling the production of large ensembles for probabilistic forecasts and data assimilation at low computational cost, data-driven models offer significant potential to improve weather forecasts. By training on reanalysis of data or observations, data-driven models can circumvent the limitations of NWP models, including biases in convection parameterization schemes that significantly affect precipitation forecasts. Once trained, data-driven models generate forecasts through orders of magnitude of inference faster than typical NWP models, allowing for the production of very large ensembles. In this context, researchers have shown that large data-driven ensembles outperform operational NWP models that can only include a limited number of ensemble members in subseasonal to seasonal (S2S) forecasts.

In addition, a considerable ensemble supports short- and long-term forecasts with data-driven predictions of extreme weather events. However, most data-driven weather models employ low-resolution data for training, often with a resolution of 5.625 or 2. Prediction of some of the broad, low-resolution atmospheric variables has been successful in last. However, the coarsening process causes the loss of important physical information at a fine scale. To be truly effective, data-driven models must provide forecasts with the same or better resolution than the most recent next-generation numerical weather models that operate at 0.1 resolution. For example, estimates with a spatial resolution of 5.625 provide a sparse grid of 32 64 pixels that represents the world.

A prediction like this cannot distinguish features smaller than 500 km. Such imprecise projections do not consider the significant impacts of small-scale dynamics on large scales and the influence of topographic factors such as mountain ranges and lakes on small-scale dynamics. As a result, low-resolution predictions can only be used in certain situations. High-resolution data (e.g., with a resolution of 0.25) can significantly improve data-driven model predictions for variables such as low-level winds (U10 and V10) that have complex fine-scale structures. , although low-resolution forecasts may be justified. for variables such as the geopotential height at 500 hPa (Z500) that do not have many small-scale structures.

Additionally, a coarser grid would not accurately represent the creation and behavior of high-impact severe events such as tropical cyclones. High-resolution models can address these issues. Their strategy: Researchers from NVIDIA Corporation, Lawrence Berkeley, Rice University, University of Michigan, California Institute of technology and Purdue University create FourCastNet, a Fourier-based neural network forecasting model, to produce global forecasts based on data from important atmospheric variables at a resolution of 0.25, or approximately 30 km near the equator, and a global grid size of 720*1440 pixels. This allows us to compare our results directly for the first time with those obtained by the ECMWF high-resolution Integrated Forecast System (IFS) model.

Figure 1 illustrates a global forecast of near-surface wind speed with a lead time of 96 hours. They highlight important high-resolution features reliably resolved and tracked by their prediction, such as Super Typhoon Mangkhut and three named cyclones (Florence, Issac, and Helene) moving toward the east coast of the United States.

In conclusion, FourCastNet offers four novel enhancements to data-driven weather forecasting:

1. FourCastNet accurately forecasts difficult variables such as surface winds and precipitation at forecast lead periods of up to one week. Surface wind forecasting on a global scale has not yet been tested using any deep learning (DL) model. Furthermore, global DL models for precipitation cannot yet resolve small-scale features. This significantly affects wind energy resource planning and disaster mitigation.

2. FourCastNet offers eight times the resolution of state-of-the-art DL-based global weather models. FourCastNet resolves severe events such as tropical cyclones and atmospheric rivers that need to be better represented by older DL models due to their coarser grids, high resolution and accuracy.

3. At lead times of up to three days, FourCastNet predictions are equivalent to those of the IFS model in terms of metrics such as root mean square error (RMSE) and anomaly correlation coefficient (ACC). Then, for lead periods up to one week, the projections of all modeled variables lag the IFS by a significant margin. FourCastNet models 20 variables at five vertical levels and is data-driven only, in contrast to the IFS model, which has been built for decades, comprises more than 150 variables at more than 50 vertical levels in the atmosphere and is governed by physics. This contrast demonstrates the immense potential of data-driven modeling to one day replace and complement NWP.

4. Compared to current PNT ensembles, which have at most about 50 members due to their high computational cost, FourCastNet’s reliable, fast and computationally affordable forecasts allow the generation of very large ensembles, allowing for the estimation of uncertainties restricted and well calibrated in extremes. with greater confidence. What is achievable in probabilistic weather forecasting is dramatically altered by the rapid development of 1,000-member ensembles, which improve the accuracy of early warnings of extreme weather events and make it possible to quickly assess their effects.

Review the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to join. our 32k+ ML SubReddit, Facebook community of more than 40,000 people, Discord channel, and Electronic newsletterwhere we share the latest news on ai research, interesting ai projects and more.

If you like our work, you’ll love our newsletter.

We are also on WhatsApp. Join our ai channel on Whatsapp.

<!– ai CONTENT END 2 –>