Predicting the weld pool and keyhole in laser beam welding through an advanced conditional latent diffusion model

Accurate prediction of the weld pool and keyhole geometries in laser beam welding (LBW) is crucial for ensuring high weld quality and optimizing process parameters. However, experimental and numerical approaches face significant challenges because of the complexity of multiphysics interactions and the associated high computational cost. In this study, we propose a novel conditional latent diffusion model (CLDM) to efficiently generate high-fidelity 2D profiles of weld pool and keyhole geometries based on welding parameters. The proposed CLDM combines a variational autoencoder for dimensionality reduction with a conditional diffusion model operating in an eight-dimensional latent space, addressing limitations such as the unstable training and mode collapse of generative adversarial networks and the over-smoothed geometries produced by conventional variational autoencoders. The model is trained using two-dimensional weld pool and keyhole profiles obtained from an experimentally validated multiphysics numerical model. The generated profiles agree well with the numerical ground truths, with correlation coefficients for weld pool length and width of 0.98 and 0.97, respectively. Once trained, the CLDM can generate multiple weld pool and keyhole profiles within minutes, demonstrating its potential as an efficient surrogate for rapid parameter screening, process optimization, and weld quality control.