Physics-Informed Neural Networks for Fast Thermal Simulation in Laser Wire Additive Manufacturing

Robotized manufacturing processes such as Wire Arc Additive Manufacturing (WAAM) and Laser Wire Additive Manufacturing (LWAM) are inherently multi-physics and governed by partial differential equations (PDEs). Classical numerical solvers (FEM, FDM, FVM) provide accuracy but are often too slow for mechatronic use cases like real-time monitoring, model-predictive control, or high-fidelity digital twins. This paper demonstrates how Physics-Informed Neural Networks (PINNs) can approximate transient thermal fields in LWAM by minimizing PDE and boundary residuals directly, eliminating the need for meshing or time stepping at inference. We provide a reproducible PINN implementation, an ONNX-exportable inference wrapper for deployment, and a companion ONNX validator that confirms faithful model export by comparing ONNX predictions against the trained PyTorch model outputs (MAE $\lt0.001^{\circ} \mathrm{C}$).