Novel subgrade soil improvement using marble dust and rice husk ash: Prediction and validation via machine learning models

The strength of the subgrade layer in pavement design is critical, since it must bear the traffic loads. The traffic load is transferred to the subgrade layer through the pavement layers, which include the sub-base, base course, and wearing course. To qualify as subgrade material under the National Highway Authority (NHA-98) regulations, the roadbed soil must fulfill specified characteristics such as the California Bearing Ratio (CBR) and the % swell ratio. When the soil does not fulfill these requirements, the soil depth up to the point where the pavement’s structural load strains are transmitted must be replaced. This is a costly and time-consuming process. In light of this, improving and enhancing the qualities of the subgrade soil by blending it with various additional materials on the site will be a better course of action. The study’s purpose is to enhance the weak subgrade soil properties by adding two unique materials in varied proportions. The soil sample is treated with different amounts of marble dust (MD) (10%, 20%, and 30%) and rice husk ash (RHA) (4%, 6%, and 8%). Testing of the control samples and design samples is included in the experimental program of the study. Material testing included grain size distribution, plasticity index (PI), compaction tests, and CBR. This study has found that the addition of (30% MD [Formula: see text] 8% RHA) has a significant effect on CBR values and percentage swell. The proportion of swell decreased by about 94.74%, dropping from 0.76% to 0.04%. Concurrently, the CBR number increased from 0.8% to 28.9%. In addition to experimental analysis, machine learning techniques are employed to predict soil behavior and validate results, using Support Vector Regressor (SVR), XGBoost Regressor, and Multilayer Perceptron Neural Network (MLP NN). The models are optimized through hyperparameter tuning using the HyperOpt library and evaluated using a fivefold cross-validation approach. Among the models, SVR emerged as the most accurate, achieving an [Formula: see text]-value exceeding 0.97, and demonstrating superior predictive performance compared to XGBoost and MLP NN. By integrating experimental methods with advanced machine learning, this research provides a novel framework for efficient and sustainable subgrade stabilization, highlighting the potential of artificial intelligence in geotechnical applications. This combined approach not only improves subgrade strength but also reduces the need for extensive soil replacement, aligning with sustainable construction practices and promoting the reuse of industrial and agricultural byproducts.