Strength analysis of sugarcane bagasse ash and rice husk based subgrade material with natural soil using different machine learning

Uniyal, Khushboo; Pratap, Bheem; Kumar, Pramod; Gupta, Keerat Kumar; Vaishnava, Rahul

doi:10.22059/ceij.2025.389294.2239

Strength analysis of sugarcane bagasse ash and rice husk based subgrade material with natural soil using different machine learning

Articles in Press

Authors

Khushboo Uniyal ¹

Bheem Pratap ¹

Pramod Kumar ²

Keerat Kumar Gupta ¹

Rahul Vaishnava ³

¹ Department of Civil Engineering, Graphic Era (Deemed to be University), Dehradun, Uttarakhand-248002, India

² 3Assistant Professor, Department of Civil Engineering, Mohan Babu University (SVEC), Tirupati, Andhra Pradesh, India-517102.

³ 5Assistant Professor, Department of Civil Engineering, Graphic Era (Deemed to be University), Dehradun, Uttarakhand-248002, India.

10.22059/ceij.2025.389294.2239

Abstract

Sugarcane bagasse ash (SBA) is a waste product that is left over after the sugar and alcohol industries. Rice husk ash (RHA) is a byproduct of rice husk burning. A soil stabilizer is terrazyme, a liquid that enhances the engineering properties of soil. In this study, natural soil, SBA and rice husk were combined in different proportions along with different concentration of terrazyme. This study uses measures like root mean square error (RMSE), mean absolute error (MAE), R2, and standard deviation to examine the performance of six predictive models across training and testing datasets: artificial neural network (ANN), XGBoost, random forest, M5P, linear regression, and non-linear regression. The ANN model exhibits overfitting as it performs poorly on unseen data (R2 = 0.589) but well on training data (R2 = 0.971). Nearly flawless training results (R2 = 0.999, RMSE = 0.005) are obtained using XGBoost, while modest generalization (testing R2 = 0.798) indicates diminished but still respectable performance on fresh data. Although Random Forest performs well during training (R2 = 0.977), there is a discernible decline in generalization (R2 = 0.853 during testing). The M5P model exhibits extreme overfitting; its testing R2 drops to 0.229 from its training R2 of 0.983.

Keywords

Sugarcane bagasse ash

Unconfined compressive strength

Artificial neural network

XGBoost and Random Forest

Subjects