Hybrid Modeling for Materials Processing by Combining Physics-Based and Data-Driven Models
Please login to view abstract download link
The exploitation of priorly-known fundamental physical rules into machine learning predictions in metallic materials mechanics and processing facilitates reduced prediction errors and enhanced generalization. Physics-based models contain assumptions and simplifications, resulting in inaccuracies, whereas data-driven models often require substantial data sets to accurately represent fundamental relationships, which can be mitigated through a synergistic integration of both modeling methodologies, respectively [1]. This proposed hybrid modeling approach consists of a physics-based model (analytical or numerical) that is refined through a data-driven discrepancy model to achieve the target solution, which originates either from experiments or high-fidelity simulations. Various complex process quantities that describe process behavior, process temperature or resulting material properties serve as prediction targets, respectively. Three application examples are presented: Laser-Shock-Peening, a residual-stress modification technique that can be used to slow down crack-growth [2]; Friction Surfacing, a solid-state processing that can be used for additive manufacturing [3]; as well as Hot Rolling, which can be used to tailor geometries and mechanical properties of metal strips [4]. Furthermore, physics-based feature engineering via normalization of inputs and outputs into dimensionless terms based on a dimensional analysis according to the Buckingham Pi theorem is evaluated with respect to a reduction of prediction scatter and the capability to perform physical extrapolation. Overall, it will be demonstrated that incorporating physics into the data science workflow can improve generalization and prediction performance, especially when dealing with limited data and aiming for physical extrapolation.
