Tom Beucler and co-authors demonstrate in this preprint , that physically rescaling the inputs and outputs of ML algorithms help them generalize to unseen climates and geographies. Janni Yuval, Paul O’Gorman and Pierre Gentine contributed to the research.