Bidirectional prediction of melt pool morphology in laser glass forming via physics-guided deep learning

Abstract Laser local forming of glass materials provides a new, non-subtractive method for manufacturing optical surfaces. The shape of the molten pool plays a crucial role in the forming accuracy. Traditional studies often rely on the rigid coverage hypothesis, ignoring the influence of surface deformation on internal temperature and flow distribution. Based on the previously proposed surface model of the molten pool, this work presents an improved simulation framework for analyzing the internal shape evolution of laser-induced molten pools, incorporating viscosity-dependent flow, the Marangoni effect, and dynamic mesh adaptation. To achieve rapid prediction between processing parameters and molten pool parameters, a bidirectional deep learning model was established. During the modeling process, the simulation data was combined with the experimental results, enhancing the data set. The maximum prediction accuracy of this model is 99.9%, with an average accuracy exceeding 99%, while the prediction accuracy using experimental data alone is 98%. In the verification test, the prediction errors for the width and depth of the molten pool were reduced by 69% and 72.3% respectively. The results show that the combination of physical information simulation and data-driven models can improve the predictive ability of laser glass forming research.