DED bead geometry and profile prediction with multimodal spatio-temporal neural networks

Abstract

Accurate prediction of bead geometry in multi-axis directed energy deposition (DED) remains challenging due to dynamic interactions between process parameters, positional dynamics, and rapid melt pool evolution. This study introduces MST-Net, a multimodal spatio-temporal neural network, alongside a comprehensive dataset of 1.2 million samples from a five-axis DED system, integrating co-axial melt pool images, control parameters, positional variables, and high-resolution point cloud labels. MST-Net employs a hierarchical architecture to fuse spatio-temporal features, achieving state-of-the-art prediction accuracy with a mean intersection over union (mIoU) of 0.93 for cross-sectional profiles and mean average precision (mAP) of 0.95 for bead dimensions. Ablation studies reveal melt pool images as the most critical input modality, while positional data governs peak localization. Temporal analysis shows asymmetric weighting of historical (look-back) and future (look-ahead) contexts optimizes predictions, with 201-step sequences balancing accuracy and computational efficiency at 116 FPS, which provides a foundation for real-time control given the predictive model. Transfer learning experiments highlight MST-Net’s adaptability, maintaining 0.95 mAP with only 10 % of training data for new printing path geometries. By addressing data scarcity and spatio-temporal complexity, this work advances predictive modeling for multi-axis DED, offering a robust framework for real-time process control in industrial applications. The dataset and model are publicly released to foster innovation in metal additive manufacturing.