Determination of the best Machine Learning model for the prediction of the California Bearing Ratio of soils in Abancay, 2024

Abstract

 The California Bearing Ratio (CBR) is a fundamental index in geotechnical engineering to evaluate the bearing capacity of soils, especially in the design and construction of pavements and other structures on natural ground. However, the determination of this index is a costly and laborious task, for that reason in this study, the prediction of CBR using machine learning models is proposed. Three machine learning models were developed, deep neural networks (DNN), decision trees, and support vector machines. The work consisted of collecting 310 records with soil characteristics, of which 217 records were considered for training, 62 for validation and 31 for testing; the data were collected in 3 soil laboratories in the city of Abancay, province of Abancay in the Apurimac region of Peru, where the following physical soil characteristics were obtained: gravel percentage, percentage of fines, optimum moisture content (OCH), liquid limit, plastic limit, plasticity index and maximum dry density (MDS) and for the characteristic to be predicted the CBR value at 100%. The models were evaluated with the coefficient of determination (R²), the mean absolute error (MAE), the mean square error (MSE), and the root mean square error (RMSE). The results show that the decision tree algorithm or model is the most efficient for predicting the CBR at 100% because it has the best coefficient of determination R² = 0.9307 and also the lowest values for the MSE = 9.199, MAE = 1.216 and RMSE = 3.033; these values are the best in relation to those found for the deep neural network and support vector regression machine models.