Abstract
This study evaluated whether data-driven machine learning models could predict the mechanical properties of friction stir-welded AA6064 joints more accurately than a calibrated physics-based finite element analysis (FEA) surrogate. The optimal parameter settings for maximizing joint efficiency were identified. Using an L9 Taguchi experimental design, 8 mm thick AA6064 plates were joined using an H13 taper-threaded tool across various rotational speeds (650–1000 rpm), traverse speeds (20–40 mm/min), axial forces (200–300 N), and a constant tool tilt angle of 2°. The resulting welds demonstrated yield strengths of 125–153 MPa, ultimate tensile strengths (UTS) of 179–216 MPa, and elongation rates of 2.19–5.90%, reaching a peak joint efficiency of 81.8% relative to that of the base material. To model these outcomes, a 4–6–3 Artificial Neural Network (ANN) and a 500-tree Random Forest Regressor were trained on the experimental dataset and benchmarked against the FEA surrogate using the Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), coefficient of determination (R2), and Leave-One-Out Cross-Validation (LOO-CV). The ANN achieved the highest precision overall, yielding a training MAPE of 0.84% and an R2 of 0.974 for UTS, whereas the random forest yielded an MAPE of 1.46% and an R2 of 0.943. Under LOO-CV, the prediction errors for UTS were 4.50% for ANN and 4.30% for Random Forest, both significantly outperforming the 9.06% error rate of the FEA model. Across broader evaluations, the ANN reduced the overall prediction error from 40.82% (FEA) to 8.97%, representing an approximate 4.6-fold improvement in the accuracy. Under the optimal Taguchi process conditions (900 rpm, 30 mm/min, 250 N axial force, and 2° tilt angle), the ANN model predicted a UTS of 211.76 MPa compared with the experimental value of 216 MPa, reflecting a prediction error of only 1.96%, whereas the FEA model exhibited an 8.89% error. The traverse speed was the primary determinant of weld performance, with a normalized Gini importance of 0.450. This was followed by the axial force and rotational speed at 0.373 and 0.177, respectively. The tool tilt angle was maintained constant and thus, was not estimable. The simultaneous variation of the traverse speed and axial force in the experimental matrix, these factors were completely aliased, rendering their individual contributions indistinguishable; therefore, they were reported collectively as 0.823. An analysis of variance confirmed the statistical significance of this aliased pair for the ultimate tensile strength (UTS) (p = 0.019) and elongation (p = 0.001), whereas the rotational speed was not significant for any response (p ≥ 0.40). Five additional algorithms, namely XGBoost, Light GBM, Cat Boost, support vector regression, and Gaussian process regression, were trained using the same leave-one-out protocol. The seven models tested, ANN and Random Forest models remained the most accurate. Within the nine-condition design examined, the proposed AI/ML framework demonstrated higher predictive accuracy at a significantly lower computational cost compared to the reduced-order physics-based surrogate used for benchmarking (25 ms for training and 149 µs per query on a single CPU core). Given that the models were trained on only nine experimental conditions and validated through leave-one-out cross-validation, y-randomization, and a single confirmation run rather than an independent test set, the reported accuracies should be considered indicative of performance within the studied process window (650–1000 rpm, 20–40 mm·min⁻¹, 200–300 N) rather than evidence of general extrapolative capability. Bootstrap 95% confidence intervals on the cross-validated errors span approximately ±1.5 %age points, and the learning curve had not saturated at eight training points, indicating that the framework is limited by data rather than by capacity.