Accuracy–Efficiency Trade-offs in Thermal Simulation of Wire-arc Directed Energy Deposition using Physics-informed Neural Networks
Please login to view abstract download link
Physics-informed neural networks (PINNs) [1] have emerged as a promising alternative to finite element methods (FEM) for thermal simulation in directed energy deposition (DED) [2–4], but most studies remain small, proof-of-concept demonstrations with disparate setups and little guidance on practical hyperparameter choices. A central barrier to adoption is the cost and uncertainty of hyperparameter tuning; in particular, the number and spatiotemporal distribution of collocation points, as well as the topology of the network, strongly affect both accuracy and training time. This work investigates the effects of network architecture and collocation-point distribution in PINNs for thermal simulation of wire-arc DED without reliance on labeled data. The transient heat equation with a moving heat source and temperature-dependent properties is enforced together with initial and mixed boundary conditions. We demonstrate accurate temperature histories and cooling-rate predictions relative to FEM baselines using a reduced network size and number of collocation points than reported in the state of the art. The results provide practical guidance for selecting key hyperparameters, advancing PINN-based DED thermal simulation from proof of concept toward more efficient configurations suitable for realistic component sizes.
