Deep learning in reduced-order space accelerates high-resolution multi-model climate downscaling

Abstract

Regional climate projections using dynamical models are computationally intensive, and generating high-resolution multi-model ensembles can be prohibitively expensive. Here, we present a lightweight deep-learning downscaling framework that generates high-resolution sea surface temperature (SST) projections within minutes on standard CPU-based systems. Our approach operates in a reduced-dimensionality space by learning relationships between principal component (PC) time series derived from coarse-resolution climate model output and high-resolution dynamically downscaled SST. By applying common empirical orthogonal function (cEOF) analysis across multiple climate models, we identify model-specific PC time series within a shared spatial basis. This allows the trained model to project future SST for independent climate models without requiring retraining for individual datasets. The proposed method demonstrates high accuracy, with the projected SST fields closely matching independent dynamically downscaled data. By constraining projections to the orthogonal EOF subspace, the framework efficiently generates high-resolution ensembles and enables robust uncertainty quantification in regional climate projections.