Performance Sensitivity of Purely Physics-Informed Neural Networks to Sampling Distribution and Architecture for Wire-arc Directed Energy Deposition

Physics-informed neural networks (PINNs) offer a clear path to scalable thermal simulation of wire-arc additive manufacturing (WAAM), also known as WADED and DED/Arc/M, by enforcing governing heat-transfer physics directly during training, thereby avoiding the need for labeled data from experiments or high-fidelity solvers. This opens the door to rapid simulation and process-parameter tuning (i.e., weld speed and temperature), necessary for structurally reliable WADED. However, the pursuit of highly accurate and efficient models is itself fraught with extensive hyperparameter tuning, while current DED PINN studies use highly variable architectures and collocation strategies, and therefore leave little indication of best practices. This work presents a controlled parametric study focused on two design choices: the collocation point sample size and distribution, and the network topology. Fully connected networks with depths 3 to 6 and widths 8 to 64 are combined with a densitybased sampling sweep using Sobol sequences, with special focus on spatiotemporal collocation points concentrated around the moving Goldak heat source. Each configuration is repeated across multiple random seeds to quantify robustness, yielding over 1500 trained models. Across all tested topologies and sampling configurations, high-accuracy solutions are achievable despite substantial variability across seeds, making reliability a key differentiator; 16-wide networks provide the best balance between computational expense, prediction accuracy, and reliability. Near-source point density typically reduces error most when the additional point density is approximately 2-3 times the base point density; however, depending on the computational budget, this benefit is diminished by the associated computational cost. The best observed model achieves an RMSE of 42.1 K in approximately 240 s of training time.