3D geometry prediction for wall deposition using transfer learning in wire arc additive manufacturing

This paper introduces a prediction model for the three-dimensional (3D) geometry of multi-layer single-track (wall-deposition) workpieces in Wire Arc Additive Manufacturing (WAAM), enabling accurate predictions with minimal experimental effort. The model extends a prior single-bead prediction model by incorporating four key enhancements: (1) using multiple cross-sections to capture the full wall geometry, (2) integration of additional input parameters to account for thermal history and deposition sequence, (3) development of an improved geometric-characterization function for better representation of wall geometry, and (4) employing a hybrid dataset composed of synthetic and experimental datasets acquired without specialized equipment, such as in-process geometry measurement systems, thereby simplifying the data collection process. A two-step transfer learning strategy was employed to pretrain the model on a synthetic dataset and subsequently train it using an experimental dataset. This approach enables accurate predictions, even when only a limited amount of experimental data is available. Compared with baseline models without transfer learning, the developed model achieved a substantial reduction in prediction errors, averaging improvements between 5–30 %. Specifically, it attained an error of approximately 10 % for height predictions and 15 % for width predictions. These contributions enhance the adaptability and scalability of the WAAM processes, thereby enabling more efficient and precise manufacturing.