Signal-driven interpretable modeling framework for the inverse design of Al/steel ultrasonic-assisted resistance spot welding

Fabricating vehicles from aluminum and steel is an effective way to balance strength with reduced weight, but the mismatch between the metallurgical and thermophysical properties of the two materials makes it difficult to achieve a reliable weld. Ultrasonic-assisted resistance spot welding (UA-RSW) has been reported to produce high-quality Al/steel joints, but the highly complex electro-thermo-mechano-acoustic system is difficult to model and thus optimize. In situ process signals from monitoring sensors offer a wealth of information on the evolution of the weld but are not directly controllable. In this study, a signal-driven interpretable modeling framework was developed that is centered on a dual-layer surrogate model. The model predicts in situ signal process features from controllable process parameters and then maps the predicted signal process features and initial process parameters to the final weld quality metrics. The model is then integrated with the non-dominated sorting genetic algorithm II to identify the Pareto front of process parameters that results in the optimal weld quality metrics. The Shapley additive explanations method is utilized to enhance the interpretability of the model. When the framework was applied to analyzing the UA-RSW process, the results indicated that ultrasonic vibrations provide a nonthermal and mechanical–metallurgical coupling effect that improves the strength of the Al/steel joint while reducing the amount of welding energy. The proposed framework offers a validated tool for transforming process signal data from monitoring sensors into actionable knowledge for the inverse design of complex manufacturing processes.