A Glass-Box Framework for Interpreting Source-Term–Related Functional Modules in a Global Deep Learning Wave Model

Abstract. Data-driven deep learning (DL) models are increasingly powerful tools for Earth system prediction, but their "black box" nature and a perceived lack of physical consistency hinder scientific trust. Validating the physical realism of these models requires new methodologies that can look inside the "black box" and map internal computations to physical processes.

This paper proposes and demonstrates such a "glass box" dissection framework. We apply this framework – which combines architectural analysis and systematic functional ablation experiments – as a case study to the OceanCastNet (OCN) v1.0 model.

The dissection demonstrates that the v1.0 model's processor autonomously learns an emergent functional partitioning. We statistically identify and validate distinct computational modules analogous to the source terms in third-generation (3G) physical wave models: a foundational propagation and climatology module (Group 4), a non-linear wind-input operator (analogous to S_in, Group 3), and a state-dependent balancing operator for dissipation (analogous to S_ds, Group 1). Furthermore, the analysis reveals that other higher-order physics are managed by a complex, coupled system of operators.

This methodological dissection provides tangible evidence of emergent physical realism in a DL model. It offers a reproducible blueprint for validating the physical fidelity of future AI-based Earth system models, providing a concrete pathway toward developing and trusting physically-constrained "grey-box" systems.