Experimental and ensemble machine learning determination of lateritic soils’ soaked CBR

Failures of flexible pavements have been attributed to the indiscriminate use of lateritic soils without prior characterisation. This study investigates the compaction behaviour of lateritic soils at Ife-Sekona Road, southwestern Nigeria. The suitability of base learners and ensemble machine-learning (ML) algorithms for predicting lateritic soils-soaked California Bearing ratio (CBR) were also tested. The Optimum Moisture Content (OMC), Maximum Dry Density (MDD) and CBR of soil samples were determined and compared using the Standard Proctor, West African Compaction and the modified AASHTO tests. Statistical comparisons were conducted using one way analysis of variance (ANOVA) and Least Significant Difference (LSD) post-hoc test at 0.05 level of significance. Base learners ML including Random Forest (RF), Elastic Net Regression (ENR), Gradient Boosted Tree (Xgboost), Support Vector Regression (SVR) and their ensemble were also employed to predict soaked CBR using 160 sample data and 10-fold cross validation. Results showed that differences in OMC and MDD between Standard Proctor and both the WACT and Modified AASHTO were significant, while differences between the latter two were not. Soaked CBR differ significantly between the three methods. The Modified AASHTO resulted in the best compaction characteristics with the highest MDD and lowest OMC. The best-performing stacked models were those combining RF and ENR, and those integrating SVM and ENR with R² = 0.80 and NSE = 0.72. MDD was the most important feature for CBR prediction from all base learners. Findings from this study provides practical guidance and data-driven decision making for improved pavement durability in tropical environments.