Sensitivity fidelity reveals true physical learning beyond forecast accuracy
Sensitivity Fidelity Reveals Physical Dependencies Beyond Forecast Accuracy in Deep-Learning Earth-System Models
Machine Learning
Summary
Deep-learning models can predict events like weather patterns well, but having a low error in predictions doesn't mean the model understands the true physical relationships involved. The authors introduced a way called sensitivity fidelity that checks how models respond to small changes in input compared to trusted references or known physics. They found models with similar prediction scores could behave very differently in these sensitivity checks. This method helps to verify if AI models learn meaningful physical dependencies and can improve trust in their forecasts.
What this means in practice
- •For climate modelers: Evaluate Earth-system models to ensure they learn physically consistent input-output relationships, not just accurate forecasts.
- •For weather forecasting teams: Improve forecast reliability by testing deep-learning models for sensitivity to input changes against known physical behavior, beyond error metrics.
Authors
Qi Chen, Peng Zhan, Hongxing Cui, Wenqing Cui, Ibrahim Hoteit
Abstract
Deep-learning models increasingly rival numerical simulation in Earth-system forecasting, yet forecast error alone cannot determine whether their learned input-output dependencies are physically consistent. Here we introduce sensitivity fidelity as an additional criterion: alongside prediction error, model responses to input perturbations should be examined against a trusted reference or established physical knowledge. In an idealized Rayleigh-Bénard convection benchmark, four surrogate architectures achieved low forecast errors yet encoded markedly different sensitivities. Quantitative comparison with a matched adjoint distinguished the models far more sharply than forecast accuracy and showed that predictive skill and sensitivity fidelity need not coincide, while finite-perturbation experiments further linked these differences to the propagation of initial-condition uncertainty. We then applied this strategy to El Niño-Southern Oscillation prediction, where no adjoint reference exists and learned sensitivities showed event-wise robustness, validity under finite perturbations, and consistency with established precursors. These results identify gradient-based evaluation as a discriminating complement to forecast-error metrics, providing a practical test of whether data-driven Earth-system models learn physically credible dependencies.