Laser-based keyhole welding is widely used in next-generation advanced manufacturing due to its high precision, speed, and efficiency. However, optimizing process parameters such as laser power, speed, and beam profile to achieve target weld geometries typically requires extensive experimental trials or computationally intensive simulations. In this work, we propose a novel Physics-Informed Neural Network (PINN) framework that enables rapid and data-efficient prediction of keyhole weld outcomes, significantly reducing the reliance on both. The approach embeds the governing heat equation and associated boundary and initial conditions directly into the neural network training, allowing the model to infer a physically consistent temperature field. By solving an inverse problem, the PINN is calibrated using limited experimental data, and once trained, it generalizes to different process conditions with negligible computational cost. Key geometric features of the melt pool—such as penetration depth and width—can then be extracted from the predicted temperature field, enabling efficient process optimization. We validate the framework on two industrially relevant scenarios with markedly different thermal and geometric characteristics: (1) a lap joint of 1.5 mm-thick EN AW-6082 aluminium alloy sheets, and (2) overlap welding of ultra-thin (0.3 mm) copper and high-purity aluminium foils, as used in electric vehicle battery tab connections. For both cases, the model achieves strong agreement with experimental results, predicting key weld features—including penetration depth and width—with errors of less than 5%. This approach allows for fast, data-efficient tuning of welding parameters, greatly reducing the time and resources needed for process development. By minimizing reliance on either extensive simulations or experiments, the methodology supports more sustainable manufacturing practices. PINN thus offers a practical and efficient tool for advancing digitalized, low-impact production aligned with Industry 4.0/5.0 goals and the shift toward net-zero manufacturing.
Piandoro et al. (Thu,) studied this question.