Predicting Resilient Modulus of Cementitiously Stabilized Subgrade Soils Using Neural Network, Support Vector Machine, and Gaussian Process Regression

Artificial neural network (ANN), support vector machine (SVM), and Gaussian process regression (GPR) were developed in this study for predicting resilient modulus (M-r) values of cementitously stabilized subgrade soils. A database was developed using laboratory test results on 160 specimens prepared by using four soils stabilized with three cementitious additives, namely, lime (3%, 6%, 9%), CFA (5%, 10%, 15%), and CKD (5%, 10%, 15%). Of these, 120 specimens (development dataset) prepared using three types of soils were used in development, and the remaining 40 specimens (validation dataset) were used in the validation of the developed models. A commercial software, MATLAB, was leveraged to develop three ANN models (radial basis function network/RBFN; a multilayer perceptrons network/MLPN with one hidden layer; and MLPN with two hidden layers); four SVM models (linear, quadratic, cubic, and radial basis function kernels); and three GPR models (rational quadratic, Matern, and exponential kernels) by using codes written in MATLAB language. The strengths and weaknesses of the developed models were examined by comparing the predicted M-r values with the observed/experimental values with respect to the mean squared error (MSE) values and determination coefficient (R-2) values. Through comprehensive comparison among these three types of models, an MLPN model with one hidden layer was determined as the best performing model developed in this study. It can be used to predict M-r of cementitiously stabilized subgrade soils for Level 2 pavement design applications. This model as well as the other models could be refined using an enriched database.

Automated summary

In this study, three machine learning models were developed to predict the resilient modulus of cementitously stabilized subgrade soils. The MLPN model with one hidden layer was determined as the best performing model among the three types of models evaluated. The developed models were compared based on mean squared error values and determination coefficient values.