Novel base predictive model of resilient modulus of compacted subgrade soils by using interpretable approaches with graphical user interface

Abstract

The flexible pavement system consists of layers that are made up of a blend of aggregates and bitumen. To design flexible pavement systems that are both safe and environmentally sustainable, it is essential to have an accurate understanding of the resilient modulus (Mr) of the compacted subgrade soil. Mr refers to the ability of the soil to resist deformation under repeated loads. Thus, a critical parameter affects the performance and longevity of pavement systems. This study employs machine learning (ML) algorithms such as individual and ensemble learners using an extensive database of 2813 data points. These include dry unit weight, weighted plasticity index, confining stress, deviator stress, moisture content, and the number of freeze-thaw cycles (FT). The individual or weak learners were incorporated to create strong and robust ensemble learners by employing techniques such as bagging, adaptive boosting, and random forest (RF). Ensemble learning methods were used to improve the performance of individual learners, such as support vector machine (SVM) and decision tree (DT), by combining their predictions. To achieve the highest R2 value, a total of twenty bagging and boosting sub-models were trained and optimized. The validation of the test data was carried out through K-Fold cross-validation, utilizing metrics such as R2, MAE, and RMSE. The developed models were rigorously tested using statistical indices (MAE, MSE, RMSE, and RMLSE) to verify their predictive accuracy, reliability, and trustworthiness. The findings indicate that the integration of bagging and boosting techniques improves the efficiency of individual machine learning (ML) models. The combination of RF and DT utilizing bagging resulted in the most reliable performance, achieving an R2 value of 0.9 and demonstrating minimum errors. In general, the implementation of the ensemble algorithm in ML improved the overall prediction accuracy of the model. Sensitivity analysis reveals that the prediction of the resilient modulus (Mr) of the subgrade is primarily influenced by dry density, confining stress, and deviator stress. Moreover, a graphical user interface (GUI) is developed for practical implantation.