Probabilistic Subseasonal-to-Seasonal Prediction of Precipitation in E3SM

Subseasonal precipitation forecasting is an often challenging task due to the inherently variable nature of daily rainfall amounts as well as their spatial specificity. Using ensemble member data from the Energy Exascale Earth System Model version 2 (E3SMv2), this research evaluates sources of probabilistic subseasonal-to-seasonal (S2S) precipitation prediction skill using explainable AI.

Here, we make use of explainable AI to identify which features of E3SM enable higher predictive skill on precipitation in the western US within S2S timeframes, and which may benefit from improved model physics. Specifically, we train neural networks to predict four-parameter conditional distributions of weekly precipitation to explore the uncertainty and predictability of rainfall in the western US. These networks leverage the signals of Pacific teleconnections (such as the El Niño Southern Oscillation and Madden-Julian Oscillation) as inputs so that these model features may be evaluated through comparative methods.

In particular, we compare network effectiveness and quality of teleconnection representation using various observational products. Such comparisons lead to an evaluation of areas of both appreciable and poor predictive skill using explainable AI methods. Together, these results inform recommended adjustments to E3SM physics affecting known teleconnections while contributing to the improvement of precipitation predictive skill on S2S timescales.