Architecture and collocation sampling of purely physics-informed neural networks for efficient and reliable thermal simulation of wire-arc directed energy deposition

Physics-informed neural networks (PINNs) offer a promising alternative to computationally demanding finite element (FE) simulations for thermal analysis of large-scale wire-arc directed energy deposition (WA-DED) without requiring labeled data. However, practical application remains hindered by a lack of systematic guidance on network architecture and collocation-point sampling, which strongly influence accuracy, robustness, and computational cost. This study presents a controlled parametric investigation of the two key PINN design choices: (i) the collocation point sample size and distribution, and (ii) the network topology. The thermal problem is governed by the transient heat conduction equation incorporating Goldak's heat source, convection, radiation, and Dirichlet boundary conditions. Fully connected networks (depths of 3–6, widths 8–64) are evaluated using a Sobol-based sampling strategy with locally increased collocation points around the moving heat source. Training across seven random seeds yielded over 1500 trained models, evaluated against a validated FE model. All investigated topologies were sufficiently expressive to achieve RMSE values below 100 K, although variability occurred across sampling configurations and seeds. Increasing near-source collocation density to 2–5 times the background density generally improves accuracy, whereas lower ratios consistently deteriorate performance. When accuracy and computational cost are considered jointly, near-source sampling is not consistently advantageous. 16-wide Networks provide the most favorable balance between accuracy, reliability, and computational efficiency. The best model achieves an RMSE of 42.1 K using 21,355 collocation points in 240 s. These findings provide actionable guidance for PINN design and establish a basis for extending WA-DED thermal simulations toward digital-twin applications.