Comparison of deep learning and physics-aware surrogate models for melt pool temperature prediction in laser welding

Abstract During laser welding, temperature changes rapidly as the keyhole and melt pool develop. Computational fluid dynamics (CFD) can simulate this behavior, but evaluating many laser power and welding speed combinations is computationally expensive. Using eight FLOW-3D simulations of AA1050 laser welding, resampled to 50 time steps, this study compares two surrogate models with different temperature-field representations. The direct deep learning surrogate predicts temperature at individual spatial locations using a temporal convolutional network-bidirectional long short-term memory (TCN-BiLSTM). Their predictions reconstruct the complete field, while a hot-zone-weighted loss emphasizes high-temperature regions. The physics-aware surrogate applies sliding-window dynamic mode decomposition (SW-DMD) to separate temperature evolution into slow and fast components. Separate proper orthogonal decomposition bases represent these components, and two MLPs predict their coefficients from laser power, welding speed, and time. For the training, validation, and test conditions, respectively, the physics-aware surrogate reduced mean RMSE from 155.4 to 39.4 K (74.6%), from 97.6 to 47.8 K (51.0%), and from 343.2 to 106.0 K (69.1%). Spatial agreement was assessed using mean intersection over union (IoU) at 900 K across three common time steps. For validation, IoU increased from 61.0 to 74.3%. However, it decreased from 79.1 to 74.1% for training and from 17.9 to 16.2% for testing. Thus, the physics-aware representation substantially reduced overall temperature error but did not consistently improve the shape and position of the hottest region. These findings distinguish improved field accuracy from reliable prediction of critical high-temperature spatial boundaries.