Abstract Achieving stable and repeatable keyhole laser welding requires precise control of process parameters, as small variations in process parameters can lead to insufficient penetration, over-penetration, or process instabilities, particularly in high-value assemblies where a single defective weld may result in the rejection of the entire component. Numerical simulation plays a central role in process understanding and optimization; however, state-of-the-art multiphysics models are characterized by high computational costs, which prevent their direct use as predictive tools for online process control, despite the growing demand for models that can be integrated within real-time monitoring and control frameworks. To overcome these limitations, a Physics-Informed Neural Network (PINN) framework is proposed as a near-real-time surrogate model for keyhole laser welding. The approach embeds the transient heat conduction equation, coupled with a double-conical volumetric heat source, directly into the neural network loss function, avoiding the need for large labelled datasets. The model is calibrated through an inverse analysis using a limited set of experiments, establishing empirical correlations between laser power, scanning speed, and heat source geometry. Validation against experimental data and high-fidelity Computational Fluid Dynamics (CFD) simulations shows good agreement, with relative errors typically below 10% for weld depth and width. Once trained, the PINN predicts the thermal field and weld bead geometry within milliseconds, enabling rapid mapping of the process window and supporting laser welding process optimization.
Piandoro et al. (Wed,) studied this question.