Optimizing Weld Quality in High Stacking Ratio Automotive Joints: Integrated Experimental Design and Machine Learning Benchmarking with Limited Datasets

Abstract

Ensuring crashworthiness in automotive body-in-white (BIW) structures requires reliable resistance spot welds meeting AWS D8.1 guidelines, which mandate minimum 20% nugget penetration into thin sheets. However, this conventional criterion based solely on nugget penetration is inadequate for high stacking ratio (HSR) joints increasingly used with advanced high-strength steels (AHSS). This research quantifies the relationship between nugget penetration and mechanical strength in dissimilar multi-sheet AHSS joints with thickness ratios ≥5:1 and develops machine learning (ML) based parameter optimization models to predict the process parameters for optimal weld joints. Systematic experimentation investigated three-sheet lap joints with thicknesses ranging from 0.65 to 2.0 mm and tensile strengths varying from 280 to 2100 MPa. A comprehensive design of experiments approach combining Box-Behnken Design (BBD) and Latin Hypercube Sampling (LHS) was implemented to optimize six welding process parameters across 80 conditions. Mechanical testing, including tensile shear strength (TSS) and cross tension strength (CTS), alongside microstructural characterization, revealed that joints without visible nugget penetration into the thin top sheet could achieve high mechanical strengths compared to fully penetrated joints. Interrupted welding experiments confirmed that bonding between sheets with high joint strength and no nugget penetration was due to either diffusion bonding or localized brazing. SEM and EDS analysis distinguished two distinct fusion interfaces: complete fusion zones with full nugget penetration and brazed interfaces, each exhibiting unique diffusion mechanisms. To extend these experimental insights, six supervised machine learning algorithms were developed and trained to predict nugget dimensions using process parameters and engineered features based on physical process relationships. Gradient boosting provided the highest predictive accuracy with R² values of 0.948 for maximum nugget width and 0.903 for nugget penetration, reducing prediction errors to 13% compared to 30% from Minitab statistical tool. Shapley additive explanation (SHAP) analysis identified welding current as the dominant process parameter, while interactions among current, weld time per pulse and electrode force proved critical for joint formation. Model-guided inverse prediction enabled dual-objective parameter optimization with experimental validation confirming predicted outcomes within target tolerances. The findings demonstrated that conventional acceptance criteria based solely on nugget penetration were inadequate for evaluating joint quality in complex dissimilar multi-sheet RSW assemblies and highlighted the need of quantitatively assessing interfacial bonding mechanisms. The validated machine learning framework provided accurate, interpretable parameter optimization, and offered a scalable pathway for broader industrial applications.