Liquefaction, which typically occurs in saturated sandy soil deposits, is one of the destructive phenomena that can occur during an earthquake. When the soil reaches liquefied state, it loses a significant amount of resistance and stiffness, which often results in widespread catastrophic damages. Therefore, accurate evaluating the potential of soil liquefaction occurrence is of great importance in earthquake geotechnical designs in regions prone to this phenomenon. The strain energy-based approach is a novel robustness technique to evaluate liquefaction potential. In the current research, 165 laboratory data sets from cyclic experiments were collected and analyzed. A predictive model using gene expression programming (GEP) was proposed to assess strain energy needed for occurrence of soil liquefaction. Assessing physical behavior of developed GEP-based model was conducted through sensitivity analysis. Performance of GEP-based was validated by comparing with a series of centrifuge experiments and cyclic triaxial tests results. Subsequently, after experimental verification of numerical modeling, the strain energy required for soil liquefaction under cyclic loading at different conditions were numerically evaluated and compared with the strain energy calculated by proposed model. Finally, the developed GEP-based model was compared with established strain energy-based relationships. The results indicated high precision of proposed GEP-based model in determination of strain energy required for soil liquefaction triggering.

Accurate prediction of resilient modulus (MR) in compacted subgrade soil is crucial for planning secure and environmentally friendly flexible pavement systems. This research assembled a dataset of 2813 data points from twelve compacted soils. The dry density, confining stress, deviator stress, number of freeze-thaw cycles, and moisture content were among the important variables considered for determining the MR. Subsequently, this study employs ensemble machine learning methodologies, specifically gene expression programming (GEP) and multi-expression programming (MEP), to investigate the subject further. The precision and anticipatory proficiency of both the GEP and MEP models are assessed through statistical evaluations, encompassing crucial metrics (R, RMSE, MAE, RSE, RRMSE, and rho). The GEP and MEP models align well with validation criteria, underscoring their robustness in predicting novel data and showcasing their broad applicability. The GEP model consistently outperformed the MEP model, with higher coefficient of regression (R2) values in both training (0.992 vs. 0.983) and testing (0.981 vs. 0.972) phases, demonstrating its superior predictive accuracy and robustness. In summary, the GEP model consistently outperforms the MEP model in accuracy and prediction, making it the preferred choice for subgrade soil MR prediction. Sensitivity analysis was done, which ranked the parameters by their influence: dry density (26.6 %), confining stress (22.7 %), weighted plasticity index (15.3 %), moisture content (13.5 %), deviator stress (12.5 %), and freeze and thaw cycles (9.4 %). This research aims to enhance the utilization of GEP and MEP in civil engineering for more accurate and efficient MR prediction, ultimately reducing time and costs.