Interactive effects of process parameters and porosity defect evaluation in thermoplastic composite resistance welding: A model-data-driven AOA-BP neural network framework

To address parameter regulation of temperature fields and strength prediction during thermoplastic composite resistance welding, a multi-physics simulation-integrated machine learning framework for cross-scale optimization is proposed. A three-dimensional transient thermal conduction model of composite resistance welding was established, validated through experimental data from thermal imaging systems and thermocouples, achieving high-fidelity simulations of welding temperature fields. This provides a robust data-driven foundation for deep learning methodologies. Building upon a Backpropagation (BP) Neural Network (ANN) model, an Archimedes Optimization Algorithm (AOA) is introduced to enable adaptive exploration of solution spaces, establishing quantitative mappings between process parameter interactions and temperature-threshold regions (representing potential defect risks) and effective weld area. This framework achieves indirect inference of optimal parameters (current: 50 A, welding time: 15 s, clamping distance: 1 mm) with high correlation ( R 2 = 0.95). Further investigations into the mechanisms of clamping distance, welding current, and time effects on interfacial/thickness-direction weld quality, particularly the evolution of temperature-threshold regions and their implications for potential defect risks, are conducted. Results demonstrate: Insufficient welding time (≤ 10 s) or low current ( ≤ 45 A) leads to increased un-welded regions and cold joints; And excessive current ( ≥ 55 A) or prolonged time ( ≥ 20 s) induces significant thermal degradation zones. This study offers critical theoretical insights for optimizing welding processes and defect mitigation in thermoplastic composites.