| Abstrakt: |
In this study, the SPT-N value of soil is modeled using five cutting-edge machine learning procedures, comprising multilayer perceptron artificial neural networks, random forests, ridge regression, support vector regressors, and extremely gradient boosting. The hyper-parameters of these algorithms are optimized utilizing the randomized search cross-validation (RSCV) algorithm. The mean average error, root mean square error, R-squared, and variance accounted for values are applied as evaluation indicators to assess the efficiency of optimized machine learning procedures on a dataset with 1113 data. The comparison shows that the RSCV approach is effective in the hyper-parameter tuning and that the optimized machine learning procedures have tremendous prospects to evaluate the SPT-N value of soils. Among the five Optimized Machine Learning Models (OMLs) used for the testing dataset, Random Forrest (RF) and Support Vector Regression (SVR) display excellent performances (R² = 0.9205 and 0.8956, respectively). The depth and CPT cone resistance are the variables that have the greatest influence on determining the SPT-N value of soil with a variable importance score of 24.06% and 23.61%, respectively. The performance of RF and SVR is compared with the existing models. It is found that the OML models such as RF and SVR outperform the existing models. [ABSTRACT FROM AUTHOR] |
