Introducing: AI-powered medium-range weather forecasting from Google DeepMind

A machine learning-based weather forecasting model from Google DeepMind leads to better, faster, and more accessible 10-day weather predictions than existing approaches, according to a new study. The model, dubbed "GraphCast," outperformed traditional systems in 90% of tested cases. It also performed well in predictions related to extreme events, for which it was not directly trained. "We believe this marks a turning point in weather forecasting," write the authors. The gold-standard approach for weather forecasting today is "numerical weather prediction (NWP)." While the accuracy of NWP models has improved drastically over the several decades, they are costly, complex, and computationally demanding, requiring increased compute resources to improve forecast accuracy. Machine learning-based weather prediction (MLWP), which can be trained from historical data, offers an alternative. Moreover, MLWP can leverage modern deep-learning hardware for efficiency.

Here, DeepMind researchers led by Remi Lam introduce GraphCast, a machine learning-based method for medium-range weather prediction trained directly from reanalysis data of past atmospheric conditions. Implemented as a neural network, it can produce an accurate 10-day forecast in under a minute on a single TPU chip. As input, GraphCast takes the two most recent states of Earth’s weather – the current time and six hours earlier – and predicts the next state of the weather six hours ahead, providing global weather prediction coverage at a roughly 0.25° latitude/longitude resolution. These predictions can be fed back into the model as inputs to generate a longer trajectory of weather states. To evaluate the forecast skill of GraphCast, Lam et al. compared GraphCast’s accuracy to that of HRES – the most accurate deterministic medium-range weather forecasting model currently used – on a large number of weather variables and lead times. They found that GraphCast was able to significantly outperform HRES on 90% of 1380 verification targets. What’s more, the platform was better at predicting severe events, including tropical cyclone tracks, atmospheric rivers (narrow regions of the atmosphere responsible for significant poleward water vapor transport), and extreme temperature anomalies, despite not being specifically trained on them. "Our approach should not be regarded as a replacement for traditional weather forecasting methods," say the authors. "Rather our work should be interpreted as evidence that MLWP is able to meet the challenges of real-world forecasting problems and has potential to complement and improve the current best methods."