Challenges and Opportunities in Using Physics-Informed Neural Networks for Adaptive Laser Welding Control

The most reported approaches for controlling the laser welding process are the application of data-driven proportional-integral-derivative (PID) controllers. These methods allow controlling the desired weld quality using direct or indirect measurements of the melt pool as feedback signals. However, these approaches fail to generalise away from this training data. With the rise of scientific machine learning, physics-informed neural networks (PINNs) for control systems are gaining popularity. The fundamental idea is to embed the governing equations of the laser welding process into the machine learning model, i.e., a specified neural network. This paper aims to critically review current developments of PINNs for real-time control of the laser welding process. The discussion will highlight the accuracy of prediction, generalisation to unseen events, and computational latency. A series of use cases with a moving Gaussian beam and a set of shaped beams will be presented to support the findings.