indrakalita/SeasonalRainbelt

MLWAM: Physics-Guided Seasonal Prediction of the West African Monsoon Rainbelt # MLWAM: Physics-Guided Seasonal Prediction of the West African Monsoon Rainbelt ## Description Welcome to the MLWAM project — a physics-guided deep learning framework for seasonal prediction of the West African Monsoon (WAM) rainbelt. Seasonal rainfall prediction over West Africa remains a major scientific and societal challenge. The region experiences strong interannual variability driven by complex interactions between sea surface temperatures, atmospheric circulation, land–atmosphere feedbacks, and large-scale climate modes. Traditional climate models often struggle to accurately capture the spatial displacement and structural evolution of the monsoon rainbelt, particularly at extended lead times. Reliable seasonal forecasts are crucial for agriculture, water resource management, flood preparedness, and food security across the region. In this project, we develop **MLWAM (Machine Learning for West African Monsoon)** — a physics-guided machine learning framework designed to forecast the spatial evolution of the rainbelt up to approximately six months ahead. Unlike purely magnitude-based prediction systems, MLWAM explicitly targets the displacement and structural dynamics of rainfall bands, reducing sensitivity to the classical double-penalty problem common in spatial verification. The framework integrates: - Climate simulation data from CESM2 - Reanalysis data from ERA5 - Satellite-derived precipitation products (IMERG) - CRPS and ACC-based skill evaluation ### Key Objectives 1. Improve seasonal prediction of West African monsoon rainbelt displacement. 2. Develop a physics-guided deep learning architecture for multi-step forecasting. 3. Compare machine learning skill against climatology and baseline approaches. 4. Quantify performance using CRPS and anomaly correlation metrics. 5. Investigate model robustness through ablation experiments. ### Forecast Configuration - Historical inputs: t−2, t−1, t. - First forecast horizon: approximately t+4 (~2 months …