<p>The thermal conductivity of backfill soils is a critical parameter for the safe design of underground thermal energy infrastructures. Bentonite-additive-based mixtures are commonly used as a backfill material due to its low hydraulic conductivity and high thermal conductivity. Although, the experimental measurement of soil thermal conductivity is cumbersome and time-consuming owing to its complex relationships with the factors such as dry density, water content, mineral composition, metric suction, organic matter, salt concentration, and particle size. Therefore, this study included laboratory measured 219 datasets using KD2-Pro dual probe of bentonite–sand and bentonite-fly ash mixed soil at different investigating parameters. These investigating parameters are compaction characteristics (i.e. porosity, dry density and degree of saturation) and physical properties (i.e. sand, clay, silt and plasticity index). Further, this study explored the tree-based machine learning models (such as decision tree, gradient boosting decision tree, random forest and extreme gradient boosting) to estimate thermal conductivity, and the results are compared with existing prediction equations. Based on the statistical performance indices (R² = 0.980, MAE = 0.023 Wm⁻¹&#xa0;K⁻¹ and RMSE = 0.057 Wm⁻¹&#xa0;K⁻¹), gradient boosting decision tree model predicts more accurate results of thermal conductivity than the other machine learning models and empirical equations. Moreover, the SHapely Additive exPlanations (SHAP) demonstrate the impact of input variables on the modelled thermal conductivity of backfill soil.</p>