Machine learning assisted mechanical design: multi objective optimization of welding parameters for robotic gripper joint

AbsractThis study presents a data driven, mechanically grounded optimization framework to improve the tensile performance of TIG welded AISI 304 stainless steel joints used in load bearing robotic gripper applications. Unlike prior machine learning–assisted welding optimization studies that emphasize metallurgical characteristics or isolated statistical metrics, the proposed approach explicitly links welding parameters to structural performance and mechanical design requirements of robotic systems. A Taguchi based experimental design consisting of ten orthogonally arranged welding trials was employed to generate the training dataset. Gaussian Process Regression (GPR) and Support Vector Regression (SVR) models were developed to predict ultimate tensile strength (UTS) and evaluated using cross validation metrics. The GPR model demonstrated superior accuracy, achieving a coefficient of determination exceeding 0.97 with low prediction error. Feature importance analysis identified welding voltage and material thickness as dominant parameters, consistent with their influence on arc stability, heat distribution, weld pool geometry and load bearing cross sectional integrity. Multi objective optimization was performed using a Multi Objective Genetic Algorithm (MOGA), where UTS was treated as the primary mechanical objective, while heat input and a distortion related surrogate response were included as competing process related objectives. The resulting Pareto optimal solutions yielded UTS values of approximately 554–580 MPa within feasible parameter limits. Model robustness was assessed through Monte Carlo–based uncertainty quantification using ±5% stochastic variation in welding parameters. Two confirmation experiments conducted outside the training dataset demonstrated agreement between predicted and experimental UTS values, with total prediction errors below 1.3%. Overall, the proposed ML-MOGA framework offers a practical optimization strategy for reliable decision making under limited experimental data conditions.