Toward complete bead morphology prediction in wire-arc directed energy deposition via in-situ point cloud processing and deep learning

Purpose Wire-arc directed energy deposition (WADED) offers advantages for fabricating large, complex aerospace components, but consistent quality control remains challenging due to its dynamic and unstable nature. Previous studies have largely focused on predicting basic bead dimensions, without fully capturing the nonlinear relationships between process parameters and full bead morphology. This study aims to model these complex interactions using data-driven approaches. Design/methodology/approach An online monitoring system was developed to collect 50,000 sets of point cloud data from 20 deposition beads using a laser profiler, alongside high-frequency electrical signals. The two data sets were temporally and spatially aligned to train two models based on a multilayer perceptron (MLP): a polynomial fitting (PF) model and a feature fitting (FF) model, targeting bead cross-sectional prediction. Findings The 6th-order PF model provided accurate cross-section representations under most conditions but underperformed when contact angles exceeded 90º. The FF model demonstrated superior performance, achieving a test set mean square error (MSE) of 0.15, an average MSE of 0.008 for the optimal sections, and an accuracy exceeding 90% in predicting bead height and width. Originality/value This study introduces a nondestructive approach for predicting the single bead morphology in WADED by integrating process parameters with high-frequency fluctuating electrical signals. Unlike conventional heat-input-based or dimension-limited models that estimate only bead height and width, the proposed FF model directly predicts the near-complete cross-sectional morphology represented by 350 pairs of two-dimensional coordinates. By combining the experimental dataset with a multi-point FF strategy using a MLP, the approach substantially enhances prediction robustness and provides a viable framework for real-time implementation in digital twin systems. Moreover, this method offers a scalable and cost-efficient solution for industrial in-situ quality monitoring and process optimization in large-scale additive and remanufacturing applications.