Explainable ensemble machine learning and multi-objective optimization for simultaneous prediction of surface roughness, tensile strength, and elongation in fused deposition modeling

The quality of fused deposition modeling (FDM) parts is governed by a large set of interacting process parameters, and the simultaneous control of geometric quality and mechanical performance remains a central challenge in additive manufacturing (AM). This study proposes a unified, three-stage data-driven framework — Prediction → Interpretability → Optimization — applied to an experimental FDM dataset comprising 50 samples, each described by nine process parameters (layer height, wall thickness, infill density, infill pattern, nozzle temperature, bed temperature, print speed, material type, and fan speed). In the prediction stage, nine multi-output regression models (Decision Tree, Random Forest, Extra Trees, Gradient Boosting, XGBoost, LightGBM, CatBoost, Multilayer Perceptron and Stacking ensemble) were trained to predict surface roughness, tensile strength, and elongation simultaneously. Given the limited sample size, models were validated using Leave-One-Out Cross-Validation (LOOCV), five-fold cross-validation, and out-of-bag bootstrap 95% confidence intervals, rather than by a single holdout split. CatBoost achieved the best generalization, with an average LOOCV coefficient of determination of 0.797 across the three outputs (roughness R 2 = 0.940; tensile strength R 2 = 0.707; elongation R 2 = 0.744) and a bootstrap mean of 0.661 (95% CI [0.45, 0.80]). In the interpretability stage, SHapley Additive exPlanations (SHAP),permutation importance, and partial-dependence and individual-conditional-expectation analyses revealed that layer height dominates surface roughness; nozzle temperature and material type are the principal drivers of tensile strength and elongation; infill pattern is statistically non-influential. In the optimization stage, a Non-dominated Sorting Genetic Algorithm II (NSGA-II) operating on CatBoost surrogate models generated a Pareto front comprising 120 non-dominated print-parameter combinations, thereby quantifying the intrinsic trade-off between smooth surfaces and high strength. The framework provides transparent, physically consistent and directly actionable guidance for FDM process design.