Abstract
<jats:p>Characterizing subgrade soils in terms of resilient modulus (MR) is crucial for pavement design. However, the process can be expensive and time-consuming, leading to the need for more efficient alternatives. This research investigates the effectiveness of using machine learning systems to predict the resilient modulus of cohesive subgrade soils. A dataset of 400 resilient modulus measurements obtained from in-situ cone penetration tests waere collected, along with an additional 200 prediction data points from past research. The combined dataset were then divided into training and testing subsets and used to develop two machine learning predicting systems: Support Vector Machine (SVM) and Back Propagation Neural Network (BPNN). The performance of these systems was evaluated using the sum of residuals (R) and the Mean Square Error (MSE) performance index. Among the SVM models tested, the one trained with a medium Gaussian kernel demonstrated the best predictive capability. For the BPNN, the most effective configuration included four input variables, fifteen hidden neurons, and thirty-three epochs. The BPNN system outperformed the SVM system, yielding higher predictability with training and testing R values of 0.98 and 0.92, and MSE training and test values of 19.84 and 90.16, respectively. In contrast, the SVM system produced training and testing R values of 0.92 and 0.72, with MSE training and test values of 47.75 and 223.73. This research demonstrates that machine learning systems can accurately predict the resilient modulus of soils, providing acceptable error levels for pavement design and construction.</jats:p>