Coloring heat: physics-informed RGB fusion of thermal gradients for porosity prediction in laser welding

Porosity is a critical defect in laser welding that severely compromises joint performance. Existing methods primarily focus on post-weld inspection or in-process monitoring, while lacking predictive assessment and grading of porosity prior to welding, thereby limiting defect prevention and quality assurance during process planning. To address this issue, a novel physics-informed parallel deep learning framework named Temperature Gradient Fusion with parallel VGG and TCN combined with Multi-Head Self-Attention (TGF-VT-MHSA) is proposed for porosity grading prediction. First, an innovative Temperature Gradient Fusion (TGF) preprocessing module is proposed to transform temporal temperature sequences into information rich image features, allowing the deep learning model to capture both temporal dependencies and spatial gradient characteristics. Second, a parallel architecture based on Multi-Head Self-Attention (MHSA) is developed to fuse temporal and spatial features, where the multi-head mechanism computes attention weights across multiple subspaces in parallel, thereby emphasizing critical features highly correlated with porosity formation and achieving a synergistic representation of temporal dynamics and spatial gradients. Subsequently, datasets with varying oscillation amplitudes and frequencies are established through both experiments and simulations to evaluate the TGF-VT-MHSA model. Finally, analyses demonstrate that the proposed method achieves superior porosity grading performance, achieving 98.3% accuracy. Moreover, compared with traditional models, the proposed model converges faster and shows greater stability, underscoring its accuracy and reliability in porosity prevention for laser welding. The proposed model is expected to be applied to defect prediction and online quality control for different materials and welding processes, supporting real-time decision making and process optimization in intelligent manufacturing.