Interpretable machine learning for predicting the strength of 3D printed fiber-reinforced concrete (3DP-FRC)

This study aims to provide an effective and accurate machine learning approach to predict the compressive strength (CS) and flexural strength (FS) of 3D printed fiber reinforced concrete (3DPFRC). Six types of ML models were utilized in this study: random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), categorical gradient boosting (Catboost), and natural gradient boosting (NGBoost). The CS and FS data is collected from recent published papers and split into training set and testing set. The hyperparameter optimization techniques are applied to optimize the ML model parameters using a grid search strategy paired with the 5-fold cross-validation. In the testing set, XGBoost, LightGBM, Catboost, and NGBoost achieve high accuracy (R2 = 0.98, 0.98, 0.98, and 0.96, respectively) on CS prediction, which is better than that of RF and SVM (R2 = 0.90 and 0.92, respectively). High accuracy on FS prediction is also obtained in XGBoost, LightGBM, CatBoost, and NGBoost (R2 = 0.94, 0.93, 0.92, and 0.90, respectively). Furthermore, the relative importance of input variables' contribution to the mechanical performance of 3DP-FRC is disclosed via Shapley additive explanations (SHAP) analysis. The SHAP analysis identifies that water/binder ratio and ordinary Portland cement content are the most influential parameters for CS, while the loading direction and fiber volume fraction are the most significant parameters for FS. The ML models incorporated with SHAP analysis disclose the relationship between the input variables and mechanical performance of 3DP-FRC and could provide valuable information for the performance-based design of the mix proportion of 3DP-FRC.