Machine learning methods to map stabilizer effectiveness based on common soil properties

Article Type

Research Article

Publication Title

Transportation Geotechnics

Abstract

Most chemical stabilization guidelines for subgrade/base use unconfined compressive strength (UCS) of treated soils as the primary acceptance criteria for selecting optimum stabilizer in laboratory testing. Establishing optimal additive content to augment UCS involves a resource-intensive trial-and-error procedure. Also, samples collected from discrete locations for laboratory trials may not be representative of the overall site. This study aims to minimize the number of laboratory trials and help strategize sampling locations by developing spatial maps of UCS at different treatment levels for lime and cement. These spatial maps were developed using machine-learning techniques, and using a database compiled from various reported studies on lime and cement stabilization of soils in the United States. Supervised learning methods under regression and classification categories were used to quantify and classify UCS values after treatments, respectively. Commonly available soil properties like Atterberg limits, gradation, and organic contents along with treatment type and amount were used as predictors and UCS values as the response. Median R2 for the best regression model was 0.75 for lime and 0.82 for cement, while the Correct Prediction Rate (CPR) for the best classification model was 92% for lime and 80% for cement. Results showed that satisfactory predictions could be made regarding stabilizer effectiveness using simple soil information commonly available. Best performing models for cement treatment were selected for generating the spatial maps for two counties in Montana. Soil samples collected from these counties were tested with different cement contents to verify the predictions. The results indicate that the Pearson's correlation coefficient for the regression model was 0.78 and CPR for the classification model was 92%. The authors hope that this study and future studies like these will increase data-driven-decision-making in geotechnical engineering practices.

DOI

10.1016/j.trgeo.2020.100506

Publication Date

3-1-2021

Comments

Open Access, Green

This document is currently not available here.

Share

COinS