Bias-Corrected CMIP6 projections of coastal direct normal irradiance to 2050 using physics-informed multi-model downscaling

Direct normal irradiance (DNI) is the solar resource governing concentrating solar power (CSP), yet it is not a standard output of global climate models, while existing CMIP6 solar-projection studies are largely global horizontal irradiance–oriented, interior-continental, and frequently single-model or uncorrected. In this study, monthly-mean DNI for 2025–2050 at a ground-validated coastal site near Karachi, Pakistan is projected using a perfect-prognosis statistical-downscaling framework. A physics-informed transfer function regresses the DNI clearness index on the global clearness index and solar geometry, raising the driver–target correlation from 0.06 (raw GHI) to 0.99, and is applied to a 25-model CMIP6 ensemble across four Shared Socioeconomic Pathways after Quantile Delta Mapping (QDM) bias correction. Five machine-learning learners are compared under one protocol, and uncertainty is decomposed into predictive, structural, inter-model and scenario axes. The reported elastic-net projector attains held-out R 2 of 0.977 and RMSE of 7.4 kWh/m 2 /month. DNI is projected to remain near its historical mean of 146 kWh/m 2 /month, declining only 2% by 2050, with a scenario ordering consistent with an aerosol-mediated pathway acting through GHI rather than radiative forcing (highest under clean-air SSP1-2.6 and lowest under high-aerosol SSP3-7.0). Inter-model spread constitutes the dominant component of the uncertainty budget, exceeding both scenario and learner spread. Controlled ablation demonstrates that omitting bias correction induces a spurious + 8–10% brightening, reverses the trend sign, and confirms that bias correction is decision-altering. The results support the long-term bankability of tropical climates while identifying inter-model uncertainty and aerosol trajectory as the principal planning risks.