Forward Thinking, Inverse Solving: Data-Driven Weld Pool Prediction from CFD-Simulated Laser Beam Welding

Abstract

Laser beam welding is an intricate yet highly useful manufacturing method. Nevertheless, determining the right parameters for any welding task requires a rigorous understanding of the underlying physics. Dynamic beam shaping offers even more possibilities, but further expands the parameter space. To simplify the process of finding the right configuration, we introduce a machine-learning-driven approach trained on synthetic data from a high-fidelity simulation validated against experimental data. The dataset enables training a model to solve the forward problem, achieving up to 97% median accuracy on the synthetic dataset when predicting welding depth from input parameters. Using the presented forward method also helps enhance inverse predictions, such as identifying appropriate input parameters for a specific welding depth. Furthermore, we will show that both the forward and inverse models serve as useful decision-support tools, providing welding engineers with additional data to aid accurate and efficient parameter selection.