Evaluation the variability of Expert performance via Machine Learning(ML) Approach on Procedures prediction of Welding Experiments

This study presents a comprehensive evaluation of the variance between expert-based estimation procedures and machine learning (ML)–driven optimization in predicting and improving the mechanical performance of welded joints. Experimental data were integrated with three supervised ML models—Random Forest, Gradient Boosting, and a Multilayer Perceptron (MLP) to quantify predictive accuracy, analyze variance, and optimize welding process parameters. The variance analysis demonstrated strong agreement between expert measurements and ML predictions, with mean deviations typically below 5%. Among the tested models, Random Forest achieved the most consistent accuracy across hardness, tensile strength, and impact energy. Optimization using the surrogate models identified the optimal parameter window as a welding current of 86–96 A, voltage of 9.5–10.5 V, and travel speed of 70–95 mm min⁻¹, corresponding to a heat-input range of approximately 0.50–0.70 kJ mm⁻¹. Comparative analysis confirmed that ML-optimized settings replicated expert estimations while reducing experimental trial-and-error and providing clear parametric insight into process behavior. The findings establish a reproducible framework for combining expert knowledge and ML techniques to enhance welding parameter design, improve mechanical consistency, and reduce manufacturing variability.