Support vector machine-based prediction of unconfined compressive strength of Multi-Walled Carbon nanotube doped soil-fly ash mixes

This comprehensive investigation explores the profound impact of multi-walled carbon nanotubes (MWCNT) on the unconfined compressive strength of cement-fly ash reinforced clayey soil. Diverse soil-fly ash blends were meticulously formulated, incorporating varying proportions of fly ash in place of soil content. These blends were subjected to treatment with different concentrations of MWCNT and sodium hexametaphosphate (SHMP), and then unconfined compressive strength (UCS) was comprehensively evaluated for each soil-fly ash mix at varying additive concentrations. Based on UCS values, the optimal mix was identified as comprising 80% soil, 20% fly ash, 0.01% MWCNT, and 2% SHMP. Subsequently, this optimal mix was further treated with 3% cement to assess its impact on UCS. Microscopic examination revealed an average particle size of 339.45 μm for the stabilized soil mix consisting of 80% soil, 20% fly ash, 0.01% MWCNT, and 2% SHMP. Three SVM regression models were trained using radial bias function (M1), linear(M2) and polynomial (M3) kernel function respectively. The SVM model M1 trained using RBF kernel function outperformed model M1 and M2. The R^2, MAE, RMSLE, RMSE value for model M1 was found to be 0.99, 0.012, 0.008, 0.012 and 0.92, 0.082, 0.068, 0.115 under training and testing conditions respectively. The box and whisker plot for residuals and Taylor diagram also complimented the statistical results and indicated the robustness of model M1. A multiple linear regression model was also developed to compare the models developed using SVM. The R², MAE, RMSLE, RMSE value of the model trained using multiple linear regression was found to be 0.80, 0.137, 0.104, 0.187 and 0.83, 0.122, 0.096, 0.168 under training and testing conditions respectively. Overall, the SVM model developed with RBF kernel function was termed as the optimum model which can predict the UCS values with adequate accuracy. Monotonicity analysis and Sensitivity analysis also confirmed the robustness of SVM-RBF model over all other models.