Ways of physics-integration into data-driven models to represent processproperty links in materials mechanics and processing via off-the-shelf machine learning models

Abstract

Integration of physical laws into the data science workflow used to perform machine learning predictions can enable error reduction, data efficiency and enhanced generalization as well as better physical consistency. Utilizing only data may require an extensive amount to represent already-known fundamental relationships, whereas employing only physical models can consume tedious calibration time and still lead to inaccurate predictions due to inherent simplifications and assumptions [1]. In this presentation, several combinations of physics-based and data-driven models to compensate the respective disadvantages of each individual approach are demonstrated and applied to solid-state materials processing techniques and a laser-based material modification process. There are different levels of integration into the data science workflow such as physicsbased feature engineering [2], physics-assisted machine learning [3] and hybrid modelling [4]. Overall, the considertation of physics within machine learning regression tasks led to improved prediction performance, enhanced generalization while allowing for data reduction as well as further physical understanding based on model explainability.