ReMatch fixes residual bias in probabilistic downscaling for climate models

Probabilistic downscaling—predicting high-resolution weather fields from low-resolution inputs—routinely suffers from biased, under-dispersive ensembles in real-world applications. A common approach decomposes the problem into a deterministic mean predictor plus a stochastic residual generator. However, the residual distribution learned during training systematically differs from the one needed at test time due to downscaling bias. That mismatch, which the authors call "residual target misspecification," is the root cause of poor calibration—not mere predictive uncertainty miscalibration.

To close this gap, Yujin Kim, Nidhi Soma, and Sarah Dean propose ReMatch (Residual Distribution Matching). It aligns the training residual distribution toward the test-time regime using optimal transport in a low-dimensional PCA space. This preserves the statistical benefits of the mean-residual framework while reducing train-test mismatch. On controlled synthetic benchmarks with varying bias levels and a real-world HRRR–ERA5 wind field downscaling task, ReMatch substantially reduces under-dispersion, improves calibration (as measured by SSR and CRPS), and outperforms strong baselines—including the standard mean-residual model, its variants, and state-of-the-art super-resolution models like SRGAN and ESRGAN.