Using Machine Learning to Cut the Cost of Dynamical Downscaling

Global climate models (GCMs) are commonly downscaled to understand future local climate change. The high computational cost of regional climate models (RCMs) limits how many GCMs can be dynamically downscaled, restricting uncertainty assessment. While statistical downscaling is cheaper, its validity in a changing climate is unclear. We combine these approaches to build an emulator leveraging the merits of dynamical and statistical downscaling. A machine learning model is developed for each coarse grid cell to predict fine grid variables, using coarse-scale climate predictors with fine grid land characteristics. Two RCM emulators, one Multilayer Perceptron (MLP) and one Multiple Linear Regression error-reduced with Random Forest (MLR-RF), are developed to downscale daily evapotranspiration from 12.5 km (coarse-scale) to 1.5 km (fine-scale). Out-of-sample tests for the MLP and MLR-RF achieve Kling-Gupta-Efficiency of 0.86 and 0.83, correlation of 0.89 and 0.86, and coefficient of determination (R-2) of 0.78 and 0.75, with a relative bias of -6% to 5% and -5% to 4%, respectively. Using histogram match for spatial efficiency, both emulators achieve a median score of similar to 0.77. This is generally better than a common statistical downscaling method in a range of metrics. Additionally, through "spatial transitivity, " we can downscale GCMs for new regions at negligible cost and only minor performance loss. The framework offers a cheap and quick way to downscale large ensembles of GCMs. This could enable high-resolution climate projections from a larger number of global models, enabling uncertainty quantification, and so better support for resilience and adaptation planning.