Can ML-based statistical downscaling models reliably extrapolate into the future?

Providing high-resolution climate information by downscaling future climate projections from the Coupled Model Intercomparison Project (CMIP6) remains a central challenge for the regional climate modeling community. CMIP6 includes a wide range of global climate model (GCM) simulations across multiple Shared Socioeconomic Pathways (SSPs), resulting in substantial computational demand for dynamical downscaling if each member is to be fully regionalized. To address this challenge, we propose a computationally efficient statistical downscaling framework based on a U-Net architecture trained over Europe. The model learns high-resolution spatial mappings directly from reanalysis data, offering a low-cost complement to regional climate models (RCMs) for large-ensemble downscaling.

We demonstrate that the climate downscaling U-Net achieves performance comparable to the HCLIM RCM when applied to unbiased EC-Earth3-Veg simulations for both the historical period and the low-emission SSP1-2.6 scenario up to 2100. The model captures spatial temperature patterns, seasonal variability, and the amplitude of warming remarkably well in these cases, providing confidence in its ability to translate GCM-scale information into higher regional climate scales.

When the U-Net is trained exclusively on reanalysis data, its extrapolation behavior under stronger forcing scenarios becomes an important aspect to evaluate. In the high-emission SSP3-7.0 scenario, after the regional climate warms by approximately +2.0 °C beyond the conditions represented in the training data, typically during 2060-2080, the model begins to diverge modestly from the warming magnitude simulated by both the driving GCM and the HCLIM downscaling. This divergence is most pronounced during summer months, while winter temperature trends remain in close agreement. These deviations are not presented as shortcomings of the method, but rather as a clear illustration of the limits of extrapolation when statistical models are trained solely on historical climate states. Highlighting these limits is essential for understanding the robustness of statistical downscaling within and beyond the training domain, particularly for applications involving strong climate-change signals.

Finally, we investigate how the model’s capabilities evolve when future regional climate information is included in the training set. Incorporating a subset of future data markedly improves the extrapolation performance, enabling the U-Net to recover long-term warming trends and seasonal patterns consistent with HCLIM even under strong forcing. This demonstrates that the U-Net architecture can effectively learn and generalize high-resolution climate transformations when provided with an extended training domain. Overall, our findings underscore the potential of deep-learning-based downscaling for scalable, ensemble-wide applications while also clarifying the conditions under which historical-only statistical training remains reliable.