Abstract
Within the fields of materials mechanics, the consideration of physical laws in machine learning predictions can enable lower prediction errors and enhanced generalization as opposed to predictions only based on data. On the one hand, exclusive utilization of fundamental physical relationships might show significant deviations in their predictions compared to reality, due to simplifications and assumptions. On the other hand, using only data and neglecting well-established physical laws can create the need for unreasonably large data sets that can be expensive to acquire in order to contain those relationships. There are many ways to integrate prior knowledge of the involved physics into the datascience workflow to identify and utilize process-property relations. Four approaches are utilized in this thesis: experimental surrogate modeling and interpretation, physics-based feature engineering, physics-assisted data-driven modeling and explanation as well as hybrid modeling which makes use of a physics-based model that is corrected by a data-driven model. Use cases have been established for three solid-state materials processing techniques, such as Refill Friction Stir SpotWelding, Friction Riveting, and Friction Surfacing, as well as for the laser-based material modification technique of Laser Shock Peening. It is shown that any level of physical integration into a data-science workflow can enhance the prediction performance of data-driven models to various degrees and can thus assist in the closure of particular research gaps.
