Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 22: Machine Learning in Dynamics and Statistical Physics II

DY 22.3: Vortrag

Dienstag, 28. März 2023, 14:30–14:45, ZEU 160

Modelling dynamic 3D-heat transfer for laser material processing using physics-informed neural networks (PINNs) — •Michael Moeckel and Jorrit Voigt — TH Aschaffenburg, Würzburger Str. 45, 63743 Aschaffenburg

Machine learning (ML) algorithms are increasingly applied to fit complex models to empirical data and to predict on dynamical system behaviour. However, such models are not intrinsically protected from violating causality or other, well-understood physical laws. Black-box ML models offer limited interpretability. Extending ML models by including physical knowledge in the optimization procedure is known as physics-based and data-driven modelling. A promising recent development are physics informed neural networks (PINN), which ensure consistency to physical laws and measured data via appropriately designed optimization routines. Here we model the 3D time-dependent temperature profile following the passage of a laser focus at the surface of some material using PINNs. In this setting, we discuss aspects of numerically efficient training for PINNs, e.g. on a set of varying collocation points. The results from the PINN agree with finite element simulations, proving the suitability of the approach. The proposed models can be smoothly integrated in monitoring systems and naturally extend to the joint analysis of measurement data and dynamical behaviour encoded in governing equations.