https://essopenarchive.org/doi/full/10.22541/essoar.177316659.94398101/v1
*Authors: *Yue Wang, Daniele Visioni, Ben Kravitz, Douglas G MacMartin, Dhruv Balwada *10 March 2026* *Abstract* Machine learning (ML) models have shown considerable promise for the statistical downscaling of climate projections; yet, their performance can degrade under non-stationarity, when future climate statistics differ from the historical data used for training. We systematically investigate this problem by downscaling global temperature and precipitation fields from coarser to higher resolutions across multiple emission pathways and solar radiation modification (SRM) scenarios in the Max Planck Institute Earth System Model. A U-Net trained directly on historical data exhibits substantial degradation under future forcings, with the root-mean-square error (RMSE) of the global-mean temperature time series increasing from 0.43$^\circ$C to 1.32$^\circ$C as scenario warming increases. To avoid this, we propose a simple data preprocessing strategy that combines residual learning with pixel-wise linear detrending and normalization to mitigate distribution shift without architectural modifications. The resulting stationary model reduces this RMSE to below 0.27$^\circ$C across all scenarios, outperforms quantile delta mapping (QDM) in spatial and temporal metrics, and runs 5–6 times faster. Downscaling SRM scenarios presents distinct challenges compared to standard emission pathways, leading to worse performance for some commonly used downscaling methods. Finally, we present the first ML-based global downscaling of a Geoengineering Model Intercomparison Project (GeoMIP) SRM simulation, generating extra-high-resolution (0.25$^\circ$) monthly temperature projections for the G6sulfur scenario to support better regional impact assessments of SRM, as well as to provide a benchmark for further localized downscaling efforts. *Source: ESS OPEN ARCHIVE * -- You received this message because you are subscribed to the Google Groups "geoengineering" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/geoengineering/CAHJsh99A8ugpJrrFGZqOpg85tsHQo2VQZkXLRS_66eeB8rAb2w%40mail.gmail.com.
