Porosity prediction in laser beam welding with a multimodal physics-informed machine learning framework

Laser beam welding (LBW) of metallic components is a knowledge‑intensive manufacturing process whose quality depends on the complex multi‑physics. However, its engineering application is often hindered by the occurrence of porosity defects. Achieving a thorough understanding and reliable prediction of porosity defects remains difficult because it demands robust representation and reasoning over nonlinear and hard‑to‑observe physical information. In this study, we propose an integrated multimodal physics-informed machine learning (PIML) framework with the help of multi-physical modelling and experimental data to predict the porosity defects in laser beam welding of aluminum alloys. The whole framework contains a multimodal PIML model for predicting the porosity ratio and an ML-based estimator for relevant physical information. By utilizing the scalar welding parameters and high-dimensional physical information (probability of keyhole collapses, cumulative existing time of collapses, and molten pool geometry) as inputs, the multimodal PIML model shows great superiority in predicting the porosity ratio, with a reduction of the mean square error by 45%, compared with the ML model trained only with welding parameters. The ML-based estimator constructed with an encoder‐decoder architecture can accurately reproduce the critical physical information within a timeframe of seconds. By integrating these two ML models, the proposed framework advances engineering informatics by offering a scalable, physics-knowledge‑centric solution for fast and accurate porosity prediction in LBW manufacturing.