Classification of geogrid reinforcement in aggregate using machine learning techniques

Abstract

The present study proposes a novel ML methodology for differentiating between unstabilized aggregate specimens and those stabilized with triangular and rectangular aperture geogrids. This study utilizes the compiled experimental results obtained from stabilized and unstabilized specimens under repeated loading into a balanced, moderate-sized database. The efficacy of five ML models, including tree-ensemble and single-learning algorithms, in accurately identifying each specimen class was explored. Shapley’s additive explanation was used to understand the intricacies of the models and determine global feature importance ranking of the input variables. All the models could identify the unstabilized specimen with an accuracy of at least 0.9. The tree-ensemble models outperformed the single-learning models when all three classes (unstabilized specimens and specimens stabilized by triangular and rectangular aperture geogrids) were considered, with the light gradient boosting machine showing the best performance—an accuracy of 0.94 and an area under the curve score of 0.98. According to Shapley’s additive explanation, the resilient modulus and confining pressure were identified as the most important features across all models. Therefore, the proposed ML methodology may be effectively used to determine the type and presence of geogrid reinforcement in aggregates, based on a few aggregate material properties and performance under repeated loading.