Research on Transformer-Based Prediction Model for Unsupported Melt-Pool State in WAAM

Metal lattice structures have garnered significant attention in engineering applications due to their low density and excellent mechanical properties. However, traditional manufacturing methods face challenges in material utilization, forming efficiency, and dimensional accuracy. This paper proposes a Transformer-based conditional time-series prediction model to predict the unsupported melt-pool state in the Wire Arc Additive Manufacturing process, including arc length, melt-pool width, and melt-pool height. We constructed a dataset, designed the model architecture, and conducted training and performance analysis. Experimental results demonstrate that the proposed model outperforms Long Short-Term Memory in prediction accuracy and meets engineering application requirements. The Transformer model uniquely leverages self-Attention mechanisms to capture long-range dependencies in melt-pool dynamics, enabling robust nonlinear mapping of complex temporal patterns. This approach addresses the limitations of traditional sequential models in modeling WAAM's melt-pool variations, offering a novel framework for real-Time melt-pool state prediction.