Machine learning and SHAP for predicting impact strength of FDM-printed short carbon fiber-reinforced nylon

The influence of fused deposition modeling (FDM) process parameters on the Charpy impact strength of short carbon fiber-reinforced nylon (PAHT-CF) was investigated using machine learning (ML). A total of 129 specimens were fabricated by varying five key parameters: extrusion temperature, bed temperature, printing speed, layer thickness, and printing orientation. The impact strength of the samples was tested according to ISO 179. Nine ML regression models were developed and hyperparameter-optimized via grid search, randomized search, and Bayesian optimization. The coefficient of determination R 2 scores and the mean square error (MSE) values from the five-fold cross-validation were used to assess the performance of the hyperparameter settings. The prediction performance of the ML models was evaluated and compared using R 2 , MSE, root mean square error (RMSE), and mean absolute error (MAE). The decision tree model tuned with Bayesian optimization yielded the best predictive accuracy (R 2 = 0.884; lowest MSE, RMSE, and MAE). SHapley Additive exPlanations (SHAP) analysis identified printing orientation as the dominant factor affecting impact strength across all models, while nonlinear ML models effectively captured synergistic interactions among process parameters. This work demonstrates the potential of interpretable ML for predicting and explaining the mechanical properties of high-performance thermoplastic composites fabricated via FDM.