Physics-informed neural networks for 3D transient thermal simulation of MIG welding

Accurate and fast prediction of transient temperature fields in MIG/MAG welding is essential for process qualification and weld integrity, yet high fidelity mesh-based simulations can become computationally expensive for fully transient 3D problems with moving heat sources. To circumvent these computational bottlenecks, this study delineates an advanced simulation architecture founded upon Physics-Informed Neural Networks (PINNs). This paradigm shifts the solution strategy from mesh-based integration to physics-constrained optimization, ensuring thermodynamic consistency while significantly accelerating time-to-solution. The proposed architecture directly assimilates the partial differential equation governing three-dimensional transient heat conduction, incorporating a dynamic Gaussian volumetric energy source, into the network’s objective function. Consequently, the training process minimizes the residuals of the governing physics, constraining the predicted manifold to satisfy Fourier’s law of conduction without the necessity for spatial discretization. Empirical validation against previously published bead-on-plate experiments on 6061-T6 aluminum alloy substrates confirms the model’s fidelity. The framework successfully reproduces complex transient thermal profiles, yielding a minimized l 2 error metric of 2.28 K, demonstrating robust agreement with thermocouple telemetry. Furthermore, a systematic hyperparameter study was conducted, as PINN performance is highly sensitive to parameter choices. Finally, this study identified an optimal configuration that balances predictive accuracy with computational cost. Notably, the optimized hyperparameters facilitate model convergence in approximately 14 minutes utilizing a single consumer-grade graphics processing unit. This efficiency establishes the PINN framework as a viable engine for real-time digital twinning and rapid iterative design in high-throughput manufacturing environments.