Machine Learning Methods to Predict Resilient Moduli Behavior of Subgrade Soils

Due to limited budgets and complex testing procedures to determine resilient modulus (M_R), engineers often rely on correlations between modulus and other engineering properties such as unconfined compressive strength (UCS). However, a majority of such correlations are based on linear regression, which can often lead to under or over-prediction of the correlated values. With the advancement in computing techniques, it has become convenient to understand and predict the behavior of engineering materials using optimization techniques. In this study, two machine learning (ML) techniques, including Artificial Neural Network (ANN) and Random Forest (RF), were used to predict the M_R values from UCS data. Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R) were computed to assess the effectiveness of the models. The statistical analysis of the testing sets indicated that the RF model achieved a goodness of fit value of 0.97, an MAE of 12.7 MPa, and an RMSE of 18.5 MPa. Alternatively, the ANN model indicated the corresponding goodness of fit value of 0.71 with an MAE of 44.3 MPa and an RMSE of 60.9 MPa. Furthermore, among various contributing factors, UCS was identified as the primary factor in predicting M_R values for both models. Based on these findings, the RF model outperformed the ANN model in predicting unknown data within the examined parameter ranges and provides fitting parameters depending on the nature of the datasets, which avoids the overfitting effect. Therefore, this study demonstrates a progressive understanding of the potential use of the advanced computing tool to obtain more accurate resilient modulus values from the strength data.