AI for weather forecasting: How Google and DeepMind researchers approach the problem

The process of anticipating weather conditions using historical and present-day data gathered by satellites, weather stations, ships, and other means is known as "data assimilation". It is the initial section of the process.

The second is a model that predicts the evolution of variables connected to the weather through time; these models are often created using numerical weather prediction (NWP). However, due to the growing amount of meteorological data, conventional NWP-based forecasting models, which rely on computing clusters to run simulations, need to scale better. In addition, they rely on the labour- and resource-intensive input of human experts to be accurate.

A machine learning-based solution

In December 2022, DeepMind and Google introduced GraphCast, a machine-learning (ML)-based weather simulator that scales well with data and can produce 10-day forecasts in less than 60 seconds. GraphCast is superior to cutting-edge ML-based benchmarks and the world's most accurate deterministic operational medium-range weather forecasting system.

As described in their research "GraphCast: Learning Skillful Medium-Range Global Weather Forecasting," GraphCast creates an autoregressive model using graph neural networks (GNNs) in an "encode-process-decode" configuration.

According to the researchers, understanding the complex physics of fluids and other substances is perfect for GNN-based designs. Additionally, the team exploits this GNN capacity by designing a novel internal multi-mesh representation technique that enables long-distance interactions with minimum message-passing overhead.

Challenges for GraphCast

The researchers' emphasis on deterministic forecasts is a significant limitation of their study. Due to the highly nonlinear nature of weather dynamics and the inherent uncertainty of the analysis used as input to weather models, it becomes increasingly difficult to make accurate point-by-point forecasts of weather trajectories as the lead time of a forecast increases; therefore, modelling the uncertainty becomes increasingly essential.

In contrast, ensemble forecast models generate numerous forecasts from a sample of initial conditions that mirror the uncertainty of the actual initial conditions. Uncertainty is quantified using statistical measurements of the ensemble forecasts. Ensemble forecasting provides a more precise distribution estimate but requires many expensive predictions.

Conclusion

DeepMind and Google's contributions reflect a significant advancement in ML-based weather modelling and, in theory, can be applied to a far broader spectrum of environmental and other geospatial-temporal forecasting challenges.

The most important application areas are modelling various meteorological variables, seasonal and climate predictions, wildfires, deforestation, animal and human activities, etc. However, their method should be viewed as a supplement to conventional weather forecasting techniques, developed over decades, carefully tested in numerous real-world scenarios, and producing probabilistic forecasts.

Instead, their work is viewed as proof that ML-based simulation is scalable to the challenges of real-world forecasting problems and has the potential to complement and transform the finest methods in use. This research contributes to realizing the promise of ML-based simulation in the physical sciences by training on complex, real-world data and outperforming standard numerical methods.