Neural General Circulation Model - differentiable atmospheric model for weather and climate predictions

Recent advancements in machine learning (ML) have led to various data-driven approaches aimed at enhancing weather prediction and climate modeling. Typically, an ML-only method is used to improve weather prediction. While current state-of-the-art ML approaches achieve lower errors at medium-range lead times, physics-based models like ECMWF’s HRES/ENS demonstrate superior physical consistency and better forecast accuracy at longer lead times. In the realm of climate modeling, ML is often used in a hybrid approach, where ML components replace uncertain parameterizations while still adhering to the governing equations for the scales they resolve. Despite advancements in atmospheric hybrid models, current attempts face challenges such as instability over extended periods, limited applicability to idealized scenarios, and only modest improvements on longer time scales in realistic scenarios. In this talk, I will explore the effectiveness of a novel approach to atmospheric modeling suitable for both weather forecasting and climate modeling. This approach involves the development of a differentiable atmospheric model, named NeuralGCM, which for the first time combines ML techniques with governing equations in an end-to-end training using ERA5 data. NeuralGCM competes with ML models for 1-10 day forecasts and matches the European Centre for Medium-Range Weather Forecasts ensemble prediction for 1-15 day forecasts. Over longer time scales, NeuralGCM displays emergent phenomena such as seasonal cycles, monsoon patterns, and tropical cyclone formation, while achieving comparable spatial bias to a global cloud-resolving model. Additionally, we present the first successful simulation of an AMIP-like experiment using a hybrid atmospheric model.