Deep reinforcement learning-driven ensemble framework for multi-target prediction of mechanical properties in additively manufactured parts using a small dataset

Abstract

Additive manufacturing of acrylonitrile–butadiene–styrene (ABS) by fused deposition modelling (FDM) enables rapid fabrication but exhibits strong process–property sensitivity, motivating reliable multi-property prediction. Here, an accurate and explainable machine-learning framework is presented for the simultaneous prediction of Young's modulus (E) and ultimate tensile strength (UTS) of FDM-printed ABS. A dataset of 78 experimental data points was compiled from tensile tests, with nozzle temperature (220–250 °C), raster angle (0–90^°), layer thickness (0.1–0.3 mm), printing speed (20–50 mm/s), and wall count (2–6) as inputs. Eight multi-output regressors were trained, and their hyperparameters were continuously tuned using deep reinforcement learning based on Deep Deterministic Policy Gradient (DDPG), and benchmarked against grid search, random search, and TPE-based Bayesian optimization. DDPG consistently improved generalisation; the best standalone model, XGBoost, achieved a test R² of 87.908 % with MAE of 0.269 and RMSE of 0.336. Ensemble learning was then investigated through simple averaging, weighted averaging, and stacking. The simple-average CatBoost–XGBoost ensemble delivered the highest test performance with R² of 88.735 %, MAE of 0.252, and RMSE of 0.322. Pairwise Wilcoxon signed-rank testing confirmed the statistical advantage of this ensemble over the leading alternatives. Finally, Shapley Additive Explanation SHAP analysis provided interpretability, identifying raster angle as the dominant driver for E, and wall count together with raster angle as the primary drivers for UTS, while printing speed showed the weakest influence.