In construction, achieving adequate soil compaction is essential for ensuring the strength and stability of geotechnical structures, with Optimum Water Content (OWC) being a critical parameter. Traditional laboratory methods for determining the OWC are accurate but often time-consuming and resource-intensive. This study investigates the potential of advanced machine learning methods: Random Forest (RF), Support Vector Machines (SVM), and Artificial Neural Networks with Multilayer Perceptron (ANN-MLP) to predict the OWC of soil using a curated dataset of over 214 soil samples collected from the Van Don - Mong Cai expressway construction project (Vietnam). The models were developed using input factors such as specific gravity, grain size distribution, organic content, and Atterberg limits. Among the three approaches, the RF model exhibited the best performance (R2 = 0.84, RMSE = 1.07% and MAE = 0.78%) compared with other models such as ANN (MLP) (R2 = 0.44, RMSE = 2.02% and MAE = 1.61%) and SVM (R2 = 0.63, RMSE = 1.65% and MAE = 1.17%). Partial Dependence Plot (PDP) analysis further highlighted fines content, plasticity indices, and organic matter as key influencing factors with a high impact on the predictive capability of the model. The findings demonstrated that the RF model offers an accurate and efficient tool for estimating the OWC of soil, with potential to reduce reliance on extensive laboratory testing and support faster, data-driven geotechnical decision-making.
Pham et al. (Tue,) studied this question.