HyMech: AI-driven framework for physics-informed discovery of interpretable hybrid models

When first-principles models inadequately match experimental data, hybrid models integrating mechanistic and data-driven components can improve the model predictions. However, traditional hybrid modeling approaches rely on uninterpretable (black-box) data-driven components, which do not reveal any insights into the origin of the observed process-model mismatch. This study proposes HyMech, an artificial intelligence-driven framework for discovering hybrid models composed of interpretable, mechanistic-like equations derived from existing first-principles models and limited data. HyMech builds upon the AI-DARWIN engine for symbolic regression (Chakraborty, Sivaram, and Venkatasubramanian. Comput. Chem. Eng., 154, 107470 (2021)), and enhances it by embedding prior process knowledge into the search space definition, thus guiding the discovery toward physically consistent solutions. Through physics-informed function and variable pools, HyMech explores model structures that refine or correct the first-principles model backbone when process-model mismatch indicates structural deficiencies. Two case studies demonstrate the ability of HyMech to diagnose process-model mismatch, discover unknown phenomena, and identify their functional forms, yielding hybrid models that capture observations while elucidating governing mechanisms. Remarkably, HyMech proves robust against noise and data scarcity.