Estimating Optimal Additive Content for Soil Stabilization Using Machine Learning Methods

Majority of geotechnical guidelines for chemical stabilization of subgrade/base materials for pavements use unconfined compressive strength (UCS) in establishing the optimal amount of additive. Laboratory determination of UCS strengths for these stabilized soils involves multiple trials by varying amount of stabilizers to achieve target strength. This process takes copious amounts of time, energy, and workforce. In addition to that, these trials are generally made on few discrete field samples which may not be representative of the overall site. Therefore, this study is aimed towards minimizing the laboratory work along with aiding in improving the sample collection strategies by using machine learning models. For this study, statistical classification was chosen to estimate optimal additive type and content. This method was used to classify whether soil will pass or fail a target strength requirement for a given amount and type of treatment. Logistic regression (LR), discriminant analysis (DA), k-nearest neighbors (KNN), and support vector machines (SVM) were used for this purpose. Commonly measured soil properties such as Atterberg limits and gradation [reported in databases such as Soil Survey Geographic Database (SSURGO)] along with treatment amount and type were chosen as predictors and, treated UCS strength as a response. Prediction accuracy was calculated using the area under the curve (AUC), correct prediction rate, true positive rate (TPR), and false positive rate (FPR). Optimal model was reported after model development using 5-fold cross-validation.