Multi-Year-to-Decadal Temperature Prediction using a Machine Learning Model-Analog Framework

Abstract

Multi-year-to-decadal climate predictions are a key tool in understanding the range of potential regional climate futures. Here, we present a framework that combines machine learning and analog forecasting for regional predictions on these timescales. A neural network is used to learn a mask of weights that highlights important global precursors to the evolution of a specific prediction target (region, variable, and lead time). A library of mask-weighted model states, or potential analogs, are then compared to a mask-weighted observational state. The known future of the best matching potential analogs serve as the prediction for the future of the observational state. We predict 2-meter temperature using the Berkeley Earth Surface Temperature dataset for observations, with a multi-model potential analog library of CMIP6 simulations. Using a 30-year climatology reference, we compare our analog method to two other analog prediction methods and the CMIP6 library using the continuous ranked probability score to assess the quality of predicted distributions, and mean squared error to assess the quality of predicted ensemble means. For nearly all cases explored, our analog method produces skillful predictions. We find higher distribution skill over the CMIP6 library in all cases, which is due to improved prediction of the distribution spread. We find overall higher skill than other analog methods in the predicted distributions and ensemble means. Finally, we find broadly similar skill to an ensemble of bias-corrected initialized Earth system models. Benefits of our analog method include low computational cost, ensemble size flexibility, and interpretability.