Bridging the Simulation Gap: Physics-Interpretable Machine Learning Framework for In Silico Fatigue Life Prediction of Additively Manufactured Polymers

The rapid adoption of additively manufactured components in structural applications exposes a fundamental limitation of current engineering workflows: process parameter selection and performance prediction are often treated as disconnected tasks. To address this gap, this study proposes an integrated and data-efficient framework that combines statistical design of experiments, ensemble machine learning, and interpretable artificial intelligence to predict the fatigue life of fused deposition modeling (FDM) printed polylactic acid (PLA). A Plackett–Burman design was employed to efficiently explore a seven-dimensional parameter space using only 64 experimentally tested specimens. Based on this dataset, a Random Forest ensemble model achieved a coefficient of determination of R^2 = 0.996 on an independent test set. Model interpretability is ensured through SHapley Additive exPlanations (SHAP), which identify wall thickness and infill density as the dominant fatigue-governing parameters, while revealing nonlinear effects associated with nozzle temperature. Predictive uncertainty is quantified via Monte Carlo simulations, enabling the extraction of reliability-oriented metrics such as P90 and P95. With microsecond-scale inference times, the proposed framework enables real-time process optimization and supports the development of digital twins in additive manufacturing.