Physics informed neural networks for modeling heat transfer problems of laser processed materials by combining measurement data and physical knowledge

Machine learning algorithms are widely adopted in technical applications. However, they can be extended by introducing scientific knowledge in the underlying optimization problem. Physics informed neural networks (PINN) are a modelling technique combining physics and data driven approaches: ensuring consistency to physical laws and measured data. In this work, stationary and time dependent heat transfer problems for laser heated solid material are modelled by PINNs and compared with FEM solutions for two scenarios: (1) PINN exclusively trained on the underlying differential equation and conditions; (2) by integrating measured sensor data in the neural network. For the later one a higher accuracy in temperature prediction is reached. The proposed models can be integrated in monitoring systems.