Real-time penetration-state prediction in laser welding: A digital-twin architecture with Gaussian process driven Markov modeling

Real-time synchronization between physical and virtual welding is hindered by the high cost of multiphase computational fluid dynamics simulations. We propose a laser-welding digital twin (LWDT) that enables online penetration-state prediction through a Gaussian-process-parameterized Markov chain model. The LWDT integrates a physics model that supplies melt-pool and keyhole features, a data-assimilation procedure that fuses simulated keyhole areas with vision-based penetration labels, and an online-learning pipeline that continually updates transition probabilities with new process data. Initialized with 150 power–speed–thickness combinations and updated with 450 additional samples, the predictor achieves mean value error 9.48% and RMSE 9.67%, which further decrease to 4.96% and 5.14% after online learning. A 1-s sequence is generated in 0.0147 s, whereas an equivalent high-fidelity simulation requires ≈110 h, demonstrating orders-of-magnitude speedup while preserving prediction accuracy. The LWDT, thus, provides a practical route to process-level, real-time quality inference and control in laser welding.