Abstract
The utilization of physical laws in machine learning predictions within materials mechanics and processing can lead to reduced errors and increased generalization compared to only using data. The usage of only physics-based models can yield errors due to inherent simplifications and assumptions, whereas the exclusive use of data-driven models can require unreasonably large data sets to represent involved physics, which are often already well-established, and still produce physically inconsistent results. To compensate for these respective disadvantages, there are numerous methods to integrate prior physical knowledge into the data-science workflow to map process-property relationships. Three approaches are demonstrated in this presentation. First, physics-based feature engineering is performed to generate salient features based on known physical relationships between initially available features. Second, simulation-assisted data-driven modeling enables data mining of a physics-based numerical model to create additional features and to enrich the initial data set via physics-based data augmentation. And third, within a hybrid model, a physics-based process model that shows errors and is corrected via a machine learning model towards the experimental target solutions. Use case applications of these methods have been implemented for Friction Riveting and Friction Surfacing. Ultimately, it is shown that the interpretation and explanation of data-driven models can lead to further understanding; thus, can help to close existing research gaps.
