Abstract
Machine learning (ML) models are successful with weather forecasting and haveshown progress in climate simulations, yet leveraging them for useful climatepredictions needs exploration. Here we show this feasibility using NeuralGeneral Circulation Model (NeuralGCM), a hybrid ML-physics atmospheric modeldeveloped by Google, for seasonal predictions of large-scale atmosphericvariability and Northern Hemisphere tropical cyclone (TC) activity. Inspired byphysical model studies, we simplify boundary conditions, assuming sea surfacetemperature (SST) and sea ice follow their climatological cycle but persistanomalies present at the initialization time. With such forcings, NeuralGCM cangenerate 100 simulation days in ~8 minutes with a single Graphics ProcessingUnit (GPU), while simulating realistic atmospheric circulation and TCclimatology patterns. This configuration yields useful seasonal predictions(July to November) for the tropical atmosphere and various TC activity metrics.Notably, the predicted and observed TC frequency in the North Atlantic and EastPacific basins are significantly correlated during 1990 to 2023 (r=~0.7),suggesting prediction skill comparable to existing physical GCMs. Despitechallenges associated with model resolution and simplified boundary forcings,the model-predicted interannual variations demonstrate significant correlationswith the observation, including the sub-basin TC tracks (p<0.1) and basin-wideaccumulated cyclone energy (p<0.01) of the North Atlantic and North Pacificbasins. These findings highlight the promise of leveraging ML models withphysical insights to model TC risks and deliver seamless weather-climatepredictions.