Development of Models for Prediction of Soil Cohesion Using Machine Learning Algorithms

Abstract

Accurate prediction of soil cohesion is crucial for the safe and economical design of geotechnical structures. This study employed five machine learning models—Artificial Neural Network (ANN), Random Forest (RF), Support Vector Regression (SVR), Gradient Boosting (GB), and Decision Tree (DT)—to predict cohesion (c) using a laboratory dataset of 233 samples. The dataset, augmented to 5000 samples using Getel, was split into 70% training and 30% testing sets. Model performance was evaluated using R-squared and Mean Squared Error (MSE). Results showed that Random Forest outperformed other models, achieving the highest R-squared score of 0.622 and the lowest MSE of 56.74, indicating excellent model fit and high predictive accuracy. Feature importance analysis revealed that plasticity, primarily influenced by Liquid Limit (LL) with an importance score of 0.879606, and Plasticity Index (PI) with an importance score of 1.441646, significantly impacts cohesion. Natural Moisture Content (NMC) also showed significant influence with a score of 0.670434. Particle Size Distribution and Specific Gravity (Gs) also contributed to the predictions. This study demonstrates the potential of machine learning models to enhance the accuracy and efficiency of soil characterization and geotechnical engineering design in predicting soil cohesion.