2023/04/01 by E. Hosseini, Pooriya Scheel, P. Scheel +5
Engineering · Physics and Astronomy · #Additive Manufacturing Materials and Processes #Model Reduction and Neural Networks #Heat Transfer Mechanisms
paper · doi:10.1016/j.cma.2023.116019
Modelling the highly localised and rapid phenomena occurring during metal additive manufacturing (MAM) processes such as the laser powder bed fusion (LPBF) demands the adoption of very fine time- and space-discretisation and therefore high computational cost for the classical simulation approaches, namely the finite element method (FEM). Particularly, when the solution is required for a range of scenarios, e.g. in sensitivity or optimisation analyses, computation costs of such simulations are not affordable. As an alternative strategy, this study explores the application of physics informed neural networks (PINNs) as a low-cost physics-based simulation approach for the thermal analysis of the LPBF process, through which reliable transient and steady-state temperature profiles for single-track LPBF depositions are achieved. An unsupervised learning strategy is employed for PINNs to parametrically solve the heat transfer equation for the LPBF process. The trained PINNs calculate the temperature profiles and the melt-pool dimensions evolving during the LPBF process for any given set of material’s thermal properties and process conditions at practically zero computational cost. The reliability of the PINNs outcomes is verified through ground-truth data generated based on several benchmark equivalent finite element simulations.