| Abstract Scope |
In the development of reliable Digital Twin systems (DT) for Tungsten Inert Gas (TIG) welding, accurate and fast prediction of thermo-geometric responses is essential. While conventional physics based thermal simulations can accurately simulate the transient heat transfer process, it is not practical to use them directly in real-time decision making, process optimization, and online quality monitoring due to their high computational costs. To overcome this drawback, this study introduces a Physics-Constrained Multi-Task Neural Network (PC-MTNN) for simultaneously predicting the peak temperature, maximum weld pool width, and the maximum HAZ width in TIG welding. A total of 318 simulation derived cases were used to train the model under different welding currents, arc voltage, travel speed, arc gap and plate thickness. A shared neural representation was used to capture the coupled thermal and geometric behavior and task-specific output heads were used to predict individual responses of the welding process. The physical consistency was directly incorporated into the model architecture by introducing the constraint of non-negative weld pool and HAZ widths as well as the constraint that the predicted HAZ width should be equal to or greater than the weld pool width.
The proposed PC-MTNN showed a good predictive performance with unseen test data. The R2 values for the model were 0.9937, 0.9804 and 0.9634 for peak temperature, weld pool width and HAZ width, respectively. The corresponding mean absolute errors were 21.87 °C, 0.216 mm, and 0.224 mm, while the root mean square errors were 36.77 °C, 0.311 mm, and 0.475 mm, respectively. The developed surrogate model is a physics-aware computational layer for the TIG welding Digital Twins that can be used to predict, monitor, and optimise the TIG welding process in real time without the need to repeatedly execute the computationally expensive numerical simulation. |