Constraining high-emission future West African monsoon with physics-weighted deep learning ensembles
The inter-model spread in future West African monsoon rainfall under high-emission scenarios remains large due to climate model biases in large-scale circulation. Here, we develop a physics-guided artificial neural network (ANN) to constrain a subset of CMIP6 precipitation projections using sea-level pressure patterns from the Sahelian monsoon ocean-pressure index (SMOPI). Model-specific ANNs are trained to learn nonlinear SMOPI-related circulation patterns and precipitation relationships under an amplitude-preserving constraint. The physical realism of these learned teleconnections is evaluated against JRA-55 reanalysis to construct a performance-based weighted ensemble. This approach reduces end-of-century inter-model spread by 20% in high-skill models, 33% in lower-skill models, and 30% across the full ensemble. It shifts the spread budget toward models reproducing observed teleconnections by up-weighting physically consistent models and down-weighting less reliable ones without collapsing ensemble diversity. This physics-weighted deep-learning architecture delivers more coherent projections and offers computationally affordable pathways to robust climate information in data-limited regions. This study introduces a physics-guided artificial neural network to constrain CMIP6 inter-model spread in West African monsoon precipitation projections, which is achieved by assigning higher weights to physically consistent models and lower weights to less reliable ones without collapsing ensemble diversity.