Hard conservation correctors can hide a degrading model when training autoregressive emulators

Abstract

AI weather and climate emulators increasingly incorporate physical principles into their formulation. One approach is to apply hard correctors that modify network outputs so that global mass, water, or energy budgets close. Prior work introduced such training-time correctors in the CREDIT framework and reported reduced precipitation bias and improved stability. Motivated by those results, we fine-tuned a global atmosphere emulator with a water-budget corrector, using the corrected prediction in the supervised loss and evaluating through post-correction budget closure. By that measure, training appeared successful. Every delivered field closed the moisture budget to machine precision. However, raw precipitation developed a growing global low bias over 18 training epochs, while the required correction increased from about 2% to roughly 24%. The cause is a scale degeneracy. A uniform change in raw precipitation amplitude is offset by a compensating change in the correction factor, leaving the corrected field, and therefore the supervised loss, unchanged. This invariance removes the restoring force on raw precipitation amplitude, allowing other training pressures to drive drift. Two changes recovered stable behavior. We supervised the pre-correction prediction and penalized its raw budget imbalance, while the hard correction remained in place for the delivered field. The required correction returned to less than 1% within the next epoch. A controlled 2x2 ablation showed that the runaway occurred only when corrected-output supervision was combined with no imbalance penalty. Exact post-correction closure therefore says little about whether the raw model has learned the budget. When a corrector removes information from the loss, the raw fields and the applied correction need to be tracked.