Shuchang Liu and Paul O’Gorman introduce CERA , a machine learning framework designed to improve the generalizability of ML-models under climate change. CERA uses an autoencoder with latent space alignment followed by a predictor to estimate moist physics processes. Trained only on control climate data with additional unlabeled warmer climate inputs, it improves predictions in a +4 K climate and outperforms both raw input and physics informed baselines. CERA captures shifts in precipitation extremes and the structure of moisture tendencies while reducing the need for manual climate invariant feature engineering, offering promise for hybrid ML physics systems and applications such as statistical downscaling.