Discovery AI for Climate and Earth Science

Overview

Numerical weather prediction (NWP) is one of the great triumphs of computational science. For six decades, operational centers like ECMWF, NOAA, and JMA have solved the primitive equations of atmospheric dynamics on ever finer grids, consuming billions of CPU hours per year. Yet even the most powerful supercomputers struggle with two persistent challenges: the computational cost of high-resolution forecasts, and the structural limitations of parameterized sub-grid physics. AI is now addressing both.

In 2023, a sequence of papers from Google DeepMind, Huawei, NVIDIA, and Caltech demonstrated that neural networks trained on reanalysis data can match or exceed the skill of operational NWP systems at medium-range weather forecasting. GraphCast, Pangu-Weather, FourCastNet, and NeuralGCM each take a different architectural approach (graph networks, 3D vision transformers, adaptive Fourier operators, hybrid learned-physics models), but they share a common result: producing 10-day global forecasts in seconds rather than hours.

This chapter covers four interconnected topics. We begin with AI weather prediction, examining the architectures and training regimes behind the leading models (Section 51.1). We then move to climate emulation and downscaling, where the goal shifts from forecasting individual weather events to reproducing the statistical properties of climate over decades, and from coarse global grids to fine regional detail (Section 51.2). Next, we tackle extreme event detection and attribution, using AI to identify rare events in observational data and to quantify the contribution of anthropogenic forcing to their probability (Section 51.3). Finally, we build a complete regional weather model by fine-tuning a pretrained architecture on a regional domain and evaluating it against ERA5 reanalysis using WeatherBench2 (Section 51.4).

These methods build on the scientific machine learning foundations of Chapter 33 (neural operators, physics-informed losses), the scientific foundation models of Chapter 27, and the scientific simulation framework of Chapter 43. They also connect to the causal inference methods of Chapter 31, which underpin the attribution science in Section 51.3.