This study explores the feasibility of training machine learning models on a large climate model ensemble to overcome the limited sample size of observational data for seasonal forecasting. The models, trained on thousands of seasons of climate model simulations, are tested on the historical observational period (1980-2020) for predicting precipitation patterns in the western United States. Results show that these machine learning models can compete with or outperform existing dynamical models, and interpretability methods reveal the relevant physical processes driving prediction skill.

Peter B Gibson, William E Chapman, Alphan Altinok, Luca Delle Monache, Michael J DeFlorio, Duane E Waliser