Predicting the West African Monsoon with a Machine Learning Emulator

Historically, devastating famines, caused by drought, have ravaged the Sahel region. Prediction of the West African Monsoon (WAM), the region’s primary precipitation catalyst, is crucial to ensure the social and economic stability of the region. While dynamical climate models are able to simulate rainfall, their resolution incurs great computational cost. We aim to emulate dynamical model predictions for Sahel precipitation during the monsoon season through supervised machine learning (ML) and decipher these predictions using explainable artificial intelligence (XAI). Previous research links the WAM to sea-surface temperatures (SSTs), the El Niño–Southern Oscillation (ENSO), and CO₂ concentration, but using SSTs to predict Sahel precipitation with ML and XAI remains largely uninvestigated. We compare regression supervised learning models, such as random forest and extreme gradient boosting (XGBoost), for predicting the quantity of Sahel precipitation during the monsoon season, and we also compare these models' performance on a classification task: predicting the associated tercile for monsoonal Sahel precipitation. The models are trained on SST and Sahel precipitation ensemble and control run outputs from the GFDL Forecast-oriented Low Ocean Resolution version of CM2.5 (FLOR) along with CO₂ concentration. For the regression and classification models respectively, we analyze the Local Interpretable Model-Agnostic Explanation (LIME) and Shapley Additive Explanation (SHAP) values of the best-performing models to determine the best predictors of Sahel rainfall.