Deep-learning models are increasingly matching numerical simulations in Earth-system forecasting, but their accuracy does not guarantee that they have learned physically meaningful relationships. The authors of a new arXiv preprint argue that forecast error alone is insufficient to judge whether a model's input-output dependencies are consistent with physical principles.

To address this, they introduce sensitivity fidelity, a concept that goes beyond standard forecast metrics. While the abstract does not provide implementation details, the idea is to evaluate whether a model's sensitivity to inputs reflects expected physical behavior.

The work highlights a growing need for diagnostics that assess not just what AI models predict, but how they arrive at those predictions. As deep-learning Earth-system models become more common, such checks could be key to building trust in their use for scientific research.