Variability of Geotechnical Parameters and Comparison of Machine Learning Algorithms for Predicting the CBR of Lateritic Gravels for road construction in Tropical Regions

Abstract This study aimed to determine the geotechnical characteristics and variability of lateritic gravels in tropical environments, and then to compare Machine Learning Algorithms for predicting the California Bearing Ratio (CBR). A total of 159 samples were collected along the Esse–Soa road section in Central Cameroon. The specific gravity, particle size distribution (PSD), plasticity, derived and compaction parameters, and CBR were determined. The studied materials consist of the < 2 mm fraction (14.6–99.4%), skeleton (0.6–85.4%), mortar (12.3–84.6%), and fines (7.3–64%). They are based on the Maximum Dry Density (1.62–2.16 g/cm 3 ) and CBR (12–94%) values, suitable for use as subgrade and subbase layers in pavements for all traffic types, and as base layers for light traffics. Data were analyzed using descriptive statistics, resulting in scattered data, characterized by low to high coefficients of variation (4–43%). Simple linear regressions (SLR) and principal component analysis were performed to predict the CBR from the other variables. No significant correlation was observed. Multiple linear and ridge regressions were performed following the SLR. They also turned out to be limited. Random forest and support vector machine regressions were also implemented. These methods, based on the same evaluation criteria as the previous learning techniques, produced successively satisfactory models, characterized by low to very low Mean Squared Error, Root Mean Squared Error, and Mean Absolute Error values, and high to very high R 2 values (0.85–0.94). PSD and derived parameters were the most suitable for predicting the CBR in the studied materials.