Abstract:
This work introduces a novel framework for physics-aware emulation that enhances the interpretability of surrogate models used in atmospheric radiative transfer. The proposed approach employs a multifidelity approach where a semi-empirical model provides an initial physics-guided estimate, which is then refined through symbolic regression. By using sparse linear techniques such as group-least absolute shrinkage and selection operator (LASSO), the method identifies a reduced set of physically meaningful features that approximate the complex behavior of deterministic radiative transfer simulations and provide insight into the underlying physical processes. The framework is evaluated in the context of atmospheric correction of hyperspectral satellite data, where it is benchmarked against traditional Gaussian process (GP) emulators. While the physics-aware emulators show a modest increase in relative error compared to standard approaches (1.9 versus 0.5), they offer a step toward more transparent and interpretable machine learning-based emulation. The results underscore the potential of blending empirical physical models with advanced machine learning techniques to create more explainable and reliable emulators for complex Earth system processes.
