Physics-constrained deep learning bias correction of CMIP6 solar radiation over Africa and its implications for solar power planning in a changing climate

This study presents a physics-constrained deep learning (PhysConDL) framework to correct systematic biases in CMIP6 surface solar radiation (Rs) simulations across Africa. Unlike conventional approaches, our framework incorporates radiative transfer physics through specialized neural networks that maintain physical consistency between atmospheric variables. Using SARAH-2.1 satellite observations as benchmark, we demonstrate remarkable improvements across 19 CMIP6 models, reducing mean absolute errors by 70–85 % continent-wide, with RMSE reductions of 65.3 % and correlation coefficients improving from below 0.90 to exceeding 0.97. Our uncertainty analysis reveals model uncertainty dominates projections throughout the 21st century (51 %), internal variability significantly affects near-term projections (49.1 %), while scenario uncertainty remains minimal (1.4–3.8 %). The framework identifies systematic radiation reductions of 0.75–1.4 W/m2 compared to raw CMIP6 outputs, with corrections exceeding −20 W/m2 in equatorial regions and pronounced seasonal asymmetry where JJA shows substantial decreases while DJF maintains relative stability. These physically consistent corrections provide a robust foundation for long-term solar energy planning across Africa's diverse climate zones, supporting the continent's renewable energy transition as installed capacity continues to grow.