Abstract
The integration of physical laws into machine learning predictions can reduce errors and improve generalization. Relying solely on data may require extensive data sets to enable representation of underlying relationships, while solely relying on physical laws can lead to inaccurate predictions due to inherent assumptions and simplifications [1]. There are several ways to combine physics and data-driven modeling to compensate respective disadvantages. Three approaches for integrating physics into data science workflows are demonstrated in this presentation: physics-based feature engineering [2,3], physics-assisted data-driven modeling [4], and hybrid discrepancy modeling [3]. These approaches are applied to solid-state materials processing techniques and a laser-based material modification process. Overall, it is shown that integrating physics into data science workflows can enhance prediction performance as well as generalization and can enable further understanding based on model explainability.
