Variable Masking for Climate Emulation in ACE

Abstract

In the recent AI Model Intercomparison Project (AIMIP) benchmark, the ACE2.1-ERA5 emulator was shown to underestimate the true near-surface warming trend both during the training period and the out-of-sample extrapolation period. In this work, we present M-ACE2S (Masked-ACE2S), a modified version of the stochastic ACE2S model, which uses variable masking during training and reduces the reliance of M-ACE2S on any single input channel. We show that M-ACE2S improves long-term near-surface air temperature prediction. We also present a novel data normalization strategy called "global mean removal", which, combined with masking, further improves the near-surface temperature forecasts. We show that M-ACE2S outperforms the ACE2S baseline on in-sample data for ERA5 while improving training stability and generalizability of near-surface temperature to unseen forcing when paired with global mean removal.