Interpretable small sample machine learning for FWTPET joint strength prediction: Consistency with classical design of experiments analysis

Data driven modelling can support manufacturing analysis when experiments are time and resourceintensive, but its reliability under severe data constraints requires careful assessment. This study presents an integrated framework combining machine-learning prediction, interpretability, classical Taguchi–ANOVA analysis, and surrogate assisted optimization for friction welding of tube to plate assemblies using an external tool (FWTPET). Nine experimental observations from an L9 orthogonal array design were analyzed using five regression models and leave one out cross validation (LOOCV), with rotational speed, pin diameter, and interference as inputs and with-hole (WH) and without-hole (WoH) joint strengths as separate responses. Second order polynomial regression achieved the best WH results, with R 2 = 0.8983, a root mean square error (RMSE) of 3.6550 MPa, and a mean absolute error (MAE) of 2.8734 MPa. Linear regression performed best for WoH, yielding R 2 = 0.7337, RMSE of 6.0642 MPa, and MAE of 4.6814 MPa. Permutation importance, SHapley Additive exPlanations (SHAP), and partial dependence and individual conditional expectation (PDP–ICE) analyses identified rotational speed as the leading parameter, consistent with Taguchi–ANOVA; however, the secondary parameter rankings differed, with a Spearman rank correlation of 0.50 between the approaches for both responses. Genetic algorithm and particle swarm optimization identified a candidate corresponding to an experimentally observed parameter combination. The principal contribution is a systematic comparison of interpretable machine learning and classical experimental design analysis within a small sample FWTPET framework. Given the nine observation dataset and the absence of independent validation, the findings remain exploratory and do not establish generalizability to untested conditions.