V3I7P20

Regression Based and Machine Learning Predictive Models for the Strength of Expansive Clay Soil Stabilized with Pulverized Ceramic Waste and Cement Kiln Dust: Development, Validation, and Sensitivity Analysis

Eleanor Chepchirchir Kimaiyo^1*, Odero J. Brian.², Mungathia Tulatia³

Abstract

The absence of validated predictive models for ceramic waste based soil stabilization systems constitutes a critical gap that impedes their engineering adoption despite demonstrated technical performance. This study develops, validates, and compares four classes of predictive models for the unconfined compressive strength (UCS) and the California Bearing Ratio (CBR) of expansive clay soil modified with pulverized ceramic waste (CW) and cement kiln dust (CKD) as an alkali activator: (i) multiple linear regression (MLR); (ii) polynomial regression (PR); (iii) stepwise regression with interaction terms (SRI); and (iv) artificial neural networks (ANN). The experimental dataset comprises 360 UCS measurements (30 mix combinations × 4 curing ages × 3 replicates) and 90 CBR measurements (30 mixes × 3 replicates at 28 days of curing), with independent variables comprising CW content (0 to 25%), CKD content (0 to 12%), and curing time (0 to 56 days). Model performance was assessed by the coefficient of determination (R²), the adjusted R², the root mean square error (RMSE), the mean absolute error (MAE), and the fivefold cross validated coefficient of determination (CVR²). Residual diagnostics included Shapiro Wilk normality testing, Breusch Pagan homoscedasticity assessment, Durbin Watson autocorrelation detection, and variance inflation factor (VIF) multicollinearity analysis. The stepwise regression model with interaction terms achieved the best balance of accuracy and parsimony for UCS (R² = 0.974; CVR² = 0.968; RMSE = 12.4 kPa) and CBR (R² = 0.961; CVR² = 0.953; RMSE = 0.86%), while the ANN model marginally outperformed all regression models on RMSE metrics (UCS RMSE = 9.1 kPa; CBR RMSE = 0.63%) at the cost of reduced interpretability. Sensitivity analysis identified CKD content as the most influential predictor (sensitivity index = 86.7% for UCS), followed by CW content (42.3%) and curing time (38.1%). The models were validated against independent test data withheld from training and against published literature benchmarks, confirming generalizability. Design charts and worked examples for mix optimization demonstrate the practical utility of the models for site specific pavement subgrade stabilization design without exhaustive laboratory testing.

Keywords:

Predictive modelling; Expansive clay; Ceramic waste; Cement kiln dust; Unconfined compressive strength; California Bearing Ratio; Multiple linear regression; Polynomial regression; Artificial neural network; Sensitivity analysis; Mix design optimization