Process‐Based Machine Learning Observationally Constrains Future Regional Warming Projections

We present the results of a novel process-based machine learning method to constrain climate model uncertainty in future regional temperature projections. Ridge-ERA5—a ridge regression model—learns coefficients to represent observed relationships between daily near-surface temperature anomalies and predictor variables from ERA5 reanalysis in Northern Hemisphere land regions. Combining the historically constrained Ridge-ERA5 coefficients with inputs from CMIP6 future projections enables a derivation of observational constraints on regional warming. Although the multi-model mean falls within the constrained range of temperatures in all tested regions, a subset of models which predict the greatest degree of warming tend to be excluded and decomposition of the constraint into predictor variable contributions suggests error-cancellation of feedbacks in some models and regions.