This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Industrial laser welding is a rapidly emerging technology in the manufacturing sector capable of producing large volumes of high-quality welds in rapid times. Extensive research in laser-based manufacturing focuses on process parameter selection, which is key to ensuring the output quality of welds. Traditionally, to define optimal process parameters, experimental trials are employed prior to manufacturing for production. While accurate, this approach incurs high resource and time costs. Alternatively, multi-physics simulation models can be employed, however, they can also be very time consuming. Novel use of AI modelling for laser welding presents the opportunity to significantly reduce the time and therefore cost of predictive simulation. However, this approach also poses challenges, for example: defining use-case specific data structures, formulating the problem accurately, and addressing the issue of insufficient data for training robust models. To address these challenges, this study proposes a practical framework that considers key process variable (KPV) definition, design of experiments (DOE) for data collection, and a process flow for closed-loop AI-driven simulation. The latter feature, allows for rapidly reconfigurable models based on changes in upstream data. The study also presents the implementation and results of a use-case where AI-driven surrogate model is implemented to replicate the simulation of microstructural changes during the welding process. We propose a comprehensive framework for AI-driven optimisation of laser welding process parameters. The approach utilises simulation data, validated experimentally, to train the proposed AI model. In this paper we show how AI can decrease computational complexity, and therefore the reduce time and cost of physics simulation models. The framework is demonstrated by a Neural Network surrogate model, which replicates the predictive capability of a complex microstructural simulation. Utilising calculated temperature profiles, which have been validated against experimentation, the new AI workflow will predict final grain density and aspect ratio. The AI model is also benchmarked against standard Machine Learning methods to evaluate performance differences and determine the most effective approach. The paper also explores the feasibility of applying Physics-Informed Neural Network (PINN) models, which incorporates partial differential equations into the learning process, enhancing its predictive capabilities, and reducing data quantity requirements This approach shows promising potential to significantly reduce the reliance on costly experiments and simulations. By exploring AI-driven modelling, we aim to establish a more efficient pathway for advancing laser welding technology. The closed-loop framework developed in this study lays a solid foundation for future work in the field, with the potential to be applied to other physics-based problems beyond laser welding.