New Method Improves Probabilistic Downscaling in Climate Models

A new technique called ReMatch (Residual Distribution Matching) has been developed to enhance probabilistic downscaling, a critical task in atmospheric science and climate modeling. Probabilistic downscaling aims to model high-resolution fields from coarse inputs, often by separating the problem into a deterministic mean prediction and a stochastic residual generation. However, this common approach frequently leads to biased and under-dispersive results in real-world scenarios due to a fundamental issue termed "residual target misspecification." ReMatch tackles this problem by aligning the training residual distribution with the distribution expected at test time. It achieves this alignment through optimal transport in a low-dimensional PCA space, effectively closing the gap between the training and testing regimes. This innovative approach preserves the statistical benefits of the mean-residual framework while significantly improving the accuracy and calibration of the stochastic generator. Evaluations on synthetic benchmarks and a real-world wind field downscaling task demonstrated that ReMatch substantially reduces under-dispersion and improves calibration, outperforming existing methods and state-of-the-art super-resolution models. The code for ReMatch is publicly available.