Speaker
Description
For machine learning regression tasks, the combination of physics-based models with data-driven models can enhance prediction performance, data efficiency and physical consistency due to the specific use of prior knowledge in the form of validated physical laws. Because a data-driven correction is assumed to be less complex when fundamental relationships are already represented in the physics-based model, this correction task also requires less data than a task to map the entire problem. Consequently, fewer samples and tests are needed when the physics-based model is effectively utilized, which results in savings of material, energy, and time. In the presented hybrid approach, a physics-based process model that exhibits a discrepancy from the experimental target solution, based on inherent assumptions and simplifications, is corrected via machine learning. This hybrid model is implemented for the solid-state process of Friction Surfacing (FS), which is a solid-state materials processing technique to produce fine grained coatings with superior corrosion and wear properties. A smoothed-particle hydrodynamics (SPH) model serves as the physics-based model and a scarce target data set based on a Box-Behnken design of experiment is used to train the data-driven correction model. The process parameters force, rotation speed and travel speed are used to prediction the geometry, i.e. thickness and width, of the deposited layer. To perform physics-based feature engineering, the Buckingham Pi theorem is used to obtain dimensionless features and to reduce model inaccuracies as well as to increase model generalization. In addition, the computational costs of the SPH model are significantly reduced by replacing it with a surrogate ML model that is then corrected instead; therefore, the hybrid model enables rapid computation of predictions that are in very good agreement with experimental measurements.
