Journal of Advances in Modeling Earth Systems Global and Planetary Change

Mitigating Non‐Stationarity in Machine Learning–Based Downscaling of Climate Projections

Yue Wang, Daniele Visioni, Ben Kravitz, Douglas G. MacMartin, Dhruv Balwada
Wiley · 2026-11-01 · DOI 10.1029/2026ms005891

Abstract

Abstract Machine learning (ML) models perform well for climate downscaling when training and testing data follow similar statistical distributions. However, their performance can degrade as future climate statistics depart from the historical training record. We systematically investigate this problem by downscaling global temperature and precipitation fields across multiple emission pathways and two solar radiation modification (SRM) scenarios in the Max Planck Institute Earth System Model, comparing four machine learning configurations with bilinear interpolation and quantile delta mapping (QDM). A baseline ML model trained on the historical period degrades substantially under future forcing, reaching a global mean temperature error of 1.6C by 2100 under SSP5‐8.5. Preprocessing the ML model's input and target to reduce non‐stationarity via detrending and residual correction keeps the global mean error within 0.2C, comparable to that of QDM. All downscaling methods compared here show substantially lower skill when evaluated on anomalies rather than total fields, suggesting that much of their skill comes from correcting the climatological mean. QDM better preserves precipitation extremes and fine‐scale variance, whereas the ML models have lower pointwise errors. Downscaling the SRM scenarios produces larger spatial temperature biases than the emission pathways. We also generate 0.25 monthly temperature projections for G6sulfur, providing a benchmark for future regional downscaling studies.

Read at publisher →