Prediction of UCS of fine-grained soil based on machine learning part 1: multivariable regression analysis, gaussian process regression, and gene expression programming

• Published in: Multiscale and Multidisciplinary Modeling, Experiments and Design Experiments and Design, 2023-06, Vol.6 (2), p.199-222
• Main Authors: Khatti, Jitendra, Grover, Kamaldeep Singh
• Format: Article
• Language: English
• Subjects: Characterization and Evaluation of Materials; Engineering; Mathematical Applications in the Physical Sciences; Mechanical Engineering; Numerical and Computational Physics; Original Paper; Simulation; Solid Mechanics
• ISSN: 2520-8160; 2520-8179
• DOI: 10.1007/s41939-022-00137-6

The present research introduces the best architecture approach and model for predicting the unconfined compressive strength (UCS) of cohesive virgin soil by comparing the multivariable regression analysis (MRA), gaussian process regression (GPR), and gene expression programming (GEP) approaches. The present research reveals the effect of the quality & quantity of the training database and the impact of the multicollinearity on the performance and overfitting of the MRA, GPR, and GEP models. The performance of the soft computing models has been measured by root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), weighted mean absolute percentage error (WMAPE), coefficient of determination ( R 2 ), correlation coefficient ( r ), variance accounted for (VAF), Nash–Sutcliffe efficiency (NS), normalized mean bias error (NMBE), Legate McCabe's Index (LMI), root mean square error to observations’ standard deviation ratio (RSR), a20-index, index of agreement (IOA), and index of scatter (IOS) statistical tools. The performance comparison of MRA, GPR, and GEP shows that GPR model MD11 has predicted UCS of soil with high performance (R = 0.9959, VAF = 99.18, NS = 0.9909, LMI = 0.1026, RSR = 0.0952, a20-index = 100.00, IOA = 0.9487 & IOS = 0.0531) and the least prediction error (RMSE = 2.4482 N/cm 2 , MAE = 1.8840 N/cm 2 , MAPE = 5.0849 N/cm 2 , WMAPE = 0.0408 N/cm 2 , NMBE = 0.1299 N/cm 2 ). In the validation, model MD11 has achieved RMSE = 3.4849 N/cm 2 , MAE = 3.1845 N/cm 2 , R  = 0.9040, R 2  = 0.8172, confidence interval of ± 5.0% by predicting UCS of lab-tested twelve soil samples, which is acceptable. This study shows that the GPR approach predicts better UCS in the presence of multicollinearity and using a small database. The sensitivity analysis illustrates that the UCS prediction of cohesive virgin soil is very highly influenced by fine content, dry unit weight, porosity, void ratio, degree of saturation, and specific gravity of soil.