Compressive strength prediction of coconut fiber reinforced concrete using PSO optimized explainable machine learning

Abstract The use of natural fibers in concrete is a sustainable strategy to improve ductility and crack resistance; however, coconut fiber can introduce nonlinear interactions among mix constituents that complicate compressive strength prediction. This study proposes an optimized and explainable machine learning framework for compressive strength prediction of coconut fiber reinforced concrete (CFRC). Six regression algorithms, support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), light gradient boosting (LGB), extreme gradient boosting (XGB), and artificial neural network (ANN) were optimized using particle swarm optimization (PSO) on 586 experimental samples with eight inputs: cement, fine aggregate (FA), natural coarse aggregate (NCA), recycled coarse aggregate (RCA), water, fiber content, fiber length, and age. A nested grouped cross-validation scheme (outer 5-fold, inner 3-fold) was adopted for robust model selection, followed by evaluation on a locked 30% test set. The best model, XGB-PSO, achieved mean cross-validation performance of R 2 = 0.963 ± 0.013, RMSE = 3.018 ± 0.646 MPa, MAPE = 7.744 ± 1.008% and demonstrated strong generalization on the locked test set (R 2 = 0.953, RMSE = 3.71 MPa, MAPE = 8.875%). Model explainability using shapley additive explanations (SHAP), partial dependence plots (PDP), accumulated local effects (ALE), and individual conditional expectation (ICE) consistently identified age and RCA as dominant predictors, supporting data driven CFRC mix design and optimization.