Using Machine Learning to Cut the Cost of Dynamical Downscaling

Global climate models (GCMs) are commonly downscaled to understand future local change. The high computational cost of regional (RCMs) limits how many GCMs can be dynamically downscaled, restricting uncertainty assessment. While statistical downscaling is cheaper, its validity in a changing unclear. We combine these approaches build an emulator leveraging the merits dynamical and downscaling. A machine learning model developed for each coarse grid cell predict fine variables, using coarse-scale predictors with land characteristics. Two RCM emulators, one Multilayer Perceptron (MLP) Multiple Linear Regression error-reduced Random Forest (MLR-RF), downscale daily evapotranspiration from 12.5 km (coarse-scale) 1.5 (fine-scale). Out-of-sample tests MLP MLR-RF achieve Kling-Gupta-Efficiency 0.86 0.83, correlation 0.89 0.86, coefficient determination (R2) 0.78 0.75, relative bias −6% 5% −5% 4%, respectively. Using histogram match spatial efficiency, both emulators median score ∼0.77. This generally better than common method range metrics. Additionally, through “spatial transitivity,” we new regions at negligible only minor performance loss. framework offers cheap quick way large ensembles GCMs. could enable high-resolution projections larger number global models, enabling quantification, so support resilience adaptation planning.