Publication Date

12-2018

Date of Final Oral Examination (Defense)

10-26-2018

Type of Culminating Activity

Thesis

Degree Title

Master of Science in Civil Engineering

Department

Civil Engineering

Supervisory Committee Chair

Bhaskar Chittoori, Ph.D.

Supervisory Committee Member

Partha S. Mukherjee, Ph.D.

Supervisory Committee Member

Mojtaba Sadegh, Ph.D.

Abstract

Unconfined compressive strength (UCS) has been widely used as one of the primary criteria for the selection of optimum type and amount of chemical stabilizer for subgrade/base stabilization. Guidelines established by various state and federal agencies aid in selecting these optimum values by recommending an initial type and amount based on a wide range of soil index properties. A significant number of laboratory trials have to be done to establish the optimum type and amount of stabilizer for a given target strength. This process takes a copious amount of time, money, and the workforce. In addition to that, the finite number of samples brought to the laboratory for characterization of chemical stabilization might not be representative of the problematic area. This study proposes the use of machine learning models to minimize the number of trials and assist in sample collection strategies by spatial mapping of predicted stabilized strength. Supervised machine-learning approaches including regression and classification were used for predicting the quantitative and categorical (pass/fail for a given threshold strength) response respectively. The parameters that didn’t have collinearity issues and are available in the Soil Survey Geographic Database (SSURGO) were chosen as input parameters for model development. An existing dataset from Australia was used to study the effectiveness of classification techniques in establishing optimum stabilizer type and amount. This analysis showed that classification methods performed well with a median correct-rate of 0.88 and median True Positive Rate (TPR) of 0.94. After this initial analysis, a database consisting of US soils and the corresponding stabilization data was compiled. Regression models using this new dataset for US soils showed comparable or better performance than regression models reported by other researchers to predict UCS values with Root Mean Square Error (RMSE) of 0.50 MPa (72.52 psi) for lime treated soils and 0.53MPa (76.87 psi) for cement treated soils. The classification model for the US soils had a median correct-rate of 0.92 and TPR of 0.94 for lime treated soils, while the same for cement treated soils were 0.80 and 0.77. The carefully chosen model input parameters (soil properties from SSURGO) in this study not only assist in arriving at an optimal type and amount of stabilizer but also help visualize the spatial distribution of UCS values for any given area within the US thereby enhancing sample collection strategies.

DOI

10.18122/td/1471/boisestate

Download

Share

COinS