| Abstract Scope |
The molten pool governs geometry, stability, and defect formation in arc welding and metal additive manufacturing, yet its modelling remains largely process-specific. High-fidelity physical solvers provide mechanistic accuracy but are too computationally expensive for online prediction and control, whereas purely data-driven models often lack stability, extrapolation capability, and cross-process generalizability. This paper investigates whether molten-pool dynamics across different processes can be represented by a shared physics-structured model rather than by separately trained process-dependent predictors.
This paper proposes a transferable latent dynamical modelling framework with two components. First, raw molten-pool images are mapped to compact latent states using a pretrained DINO encoder, while Neural ODE regularization enforces temporally smooth and physically consistent representations. This design reduces dependence on process-specific feature engineering and provides a reusable state representation across molten-pool processes. Second, a physics-structured dynamic model derived from a lumped molten-pool energy balance specifies the dynamic form while data learn the nonlinear internal functions. The model separates a fast geometric state from a slow thermal state associated with accumulated workpiece heat, showing that pool geometry alone is not Markovian and must be augmented by a hidden thermal variable. The resulting dynamics describe molten-pool evolution as a dissipative relaxation process in which geometry contracts toward an equilibrium determined by process input and accumulated heat.
The framework combines transferable visual representations with stable and interpretable physics-structured dynamics. It is validated on GTAW simulations with known ground truth, real GTAW and GMAW experiments, and the multi-source Melt-Pool-Kinetics dataset. Across these settings, the same framework is applied without process-specific redesign, demonstrating a reusable modelling strategy for molten-pool dynamics across welding and metal additive manufacturing processes. |