Data‐Driven Medium‐Range Weather Prediction With a Resnet Pretrained on Climate Simulations: A New Model for WeatherBench

Abstract

Numerical weather prediction has traditionally been based on the models that discretize the dynamical and physical equations of the atmosphere. Recently, however, the rise of deep learning has created increased interest in purely data‐driven medium‐range weather forecasting with first studies exploring the feasibility of such an approach. To accelerate progress in this area, the WeatherBench benchmark challenge was defined. Here, we train a deep residual convolutional neural network (Resnet) to predict geopotential, temperature and precipitation at 5.625° resolution up to 5 days ahead. To avoid overfitting and improve forecast skill, we pretrain the model using historical climate model output before fine‐tuning on reanalysis data. The resulting forecasts outperform previous submissions to WeatherBench and are comparable in skill to a physical baseline at similar resolution. We also analyze how the neural network creates its predictions and find that, for the case studies analyzed, the model has learned physically reasonable correlations. Finally, we perform scaling experiments to estimate the potential skill of data‐driven approaches at higher resolutions.

Sub‐Seasonal Forecasting With a Large Ensemble of Deep‐Learning Weather Prediction Models

“…This iterative approach follows the time-stepping strategy in NWP; once our model is trained, it can provide forecasts at any lead time that is a multiple of six hours. This approach may contrasted with models trained to forecast only for specific lead times, such as 6 hours, 1, 3, and 5 days in Rasp and Thuerey (2021).…”

Machine Learning Emulation of Gravity Wave Drag in Numerical Weather Forecasting

“…Seismic changes include investigations into whether machine learning can replace the whole forecasting system, either by learning from observational data (Sønderby et al., 2020 ) or atmospheric reanalysis (Rasp et al., 2020 ). Early results are promising in the area of nowcasting (Sønderby et al., 2020 ), but still lag behind classical modeling for short and medium range forecasting (Weyn et al., 2019 ) with evidence (Rasp & Thuerey, 2021 ) and arguments (Palmer, 2020 ) that there is insufficient data when moving to high resolutions. For seasonal forecasting, machine learning techniques again show promising results, e.g., forecasting El Nino sea‐surface temperatures (Dijkstra et al., 2019 ).…”

A Comparison of Data‐Driven Approaches to Build Low‐Dimensional Ocean Models

“…Transfer learning may help when training data‐driven models for the real ocean with limited observations. In this approach, the models are trained first using the historical climate data, such as CMIP model outputs, before fine tuning using observations and reanalysis data (Ham et al., 2019; Rasp & Thuerey, 2021).…”