Permeability Prediction of Soil–Rock Mixtures Using Hybrid Machine Learning Considering Matrix Composition and Rock Content

Soil–rock mixtures (SRMs) are heterogeneous geomaterials commonly encountered in geotechnical structures such as dams, slopes, and tunnels. Due to the complex interaction between fine‐grained soil matrices and coarse rock fragments, predicting the hydraulic behavior of SRMs, particularly permeability, remains a significant challenge. This study investigates the influence of different matrix compositions, namely, loamy sand, sandy loam, and silty loam, on the permeability behavior of SRMs under varying volumetric block proportions (VBPs), rock sizes (RSs), rock shapes (RSHs), and moisture conditions. Large‐scale constant‐head permeability tests were conducted on remolded SRM specimens prepared with different matrix types (MTs) and block characteristics. Experimental results indicated that permeability increases under dry optimum conditions and decreases under wet optimum conditions due to changes in pore structure and particle arrangement. Higher VBPs generally increased permeability, while larger RSs reduced flow due to increased tortuosity. To predict permeability, hybrid machine learning (ML) models combining random forest (RF) and extreme gradient boosting (XGB) with Leopard Seal (LS) and Egret Swarm (EG) optimization algorithms were developed and validated using fivefold cross‐validation. Among the developed models, the XGB‐EG hybrid model achieved the best predictive performance on the testing dataset with R 2 = 0.988798, root mean square error (RMSE) = 0.027068, mean absolute error (MAE) = 0.01481, and relative standard error (RSR) = 0.105839. SHAP (SHapley Additive exPlanations) analysis further identified VBP as the most influential parameter affecting permeability prediction. The findings demonstrate the importance of matrix composition in governing SRM hydraulic behavior and highlight the capability of hybrid ML approaches for reliable permeability prediction in heterogeneous geomaterials.