Peak Shear Strength of Discrete Fiber-Reinforced Soils Computed by Machine Learning and Metaensemble Methods

Abstract

The accuracy of prior theoretical and empirical models for predicting the shear strength of fiber-reinforced soil (FRS) is questionable because of the difficulty of using these simplified models to describe the complex mechanism of soil-fiber interaction. This study compiled a large database of available high quality triaxial and direct shear tests on FRS documented in the literature from 1983 to 2015. The database includes information on the properties of sand, fibers, soil-fiber interface, and stress parameters. After data preprocessing, data mining technologies were employed to identify factors influencing shear strength and to predict the peak friction angle of FRS. The analysis techniques included (1) classification and regression methods, i.e., linear regression (REG) analysis, classification and regression tree (CART) analysis, a generalized linear (GENLIN) model, and chi-squared automatic interaction detection (CHAID); (2) machine learners, i.e., artificial neural network (ANN) and support vector machine (SVM) and support vector regression (SVR); and (3) metaensemble models, i.e., voting, bagging, stacking, and tiering. The analytical results indicated that fiber content, fiber aspect ratio, soil friction angle, and stress parameter had major effects on FRS shear strength. The optimal model obtained after further model training, cross-validation, and testing was the Tiering SVM-(SVR/SVR) method. The correlation coefficient (R) of the prediction values with the measured values in the database was 0.89, indicating a strong association. The mean absolute percentage error (MAPE) was 3.27%, root mean square error (RMSE) was 1.98°, and mean absolute error (MAE) was 1.07°. The overall improvement in performance measures was 9.31–79.50%, which was comparable to that of other models reported in the literature. This study contributes to the domain knowledge by developing an effective artificial intelligence (AI) model for predicting the peak friction angle of FRS.