Comparative Study of Various Machine Learning Models for Estimating Standard Penetration Test-N Value

The Standard Penetration Test (SPT) is a widely used field test in geotechnical engineering to determine the consistency, strength, and other properties of soil. This paper presents an alternative approach to determine the Standard Penetration Resistance Value. A lot of times, due to budget limitations, time constraints, and other concerns, there is a tendency to discard this test. Often, N value is estimated from the adjacent site if the data is available, else it is discarded. Various studies have been carried out to determine factors such as shear velocity and angle of internal friction to estimate SPT-N value. This research is a novel approach to estimate N value of the soil with the help of various soil parameters. Here, N value of cohesionless soil is estimated using different techniques such as artificial neural networks (ANNs), XGBoost model, ElasticNet model, lasso regression model, extra trees, Bayesian linear regression, ridge regression, AdaBoost, and random forest models with the help of seven soil parameters, namely moisture content, soil composition (% of gravel, sand, silt, and clay content), bulk density, and dry density. The model is trained on a dataset of SPT-N values and corresponding soil properties. The performance of these models is evaluated using various statistical measures and compared with the existing empirical equations. The results show that the random forest and AdaBoost models have the most efficient performance with R² values of 0.62 and 0.63, respectively, while their accuracies are 84.53% and 83.61%, respectively. The proposed approach can be a useful tool for geotechnical engineers to predict the SPT-N values of soils and thus facilitate more efficient and cost-effective site investigations.