Experimental and machine learning-based estimation of the thermal diffusivity of high-plastic soil

In many residential areas, power companies are increasingly opting for underground cables for medium- and high-voltage electricity transmission to mitigate weather-related disruptions and ensure safe electricity distribution. Thermal diffusivity (TD) is a critical thermophysical parameter influencing heat transfer processes between heat-sensitive subsurface structures such as underground power cables, ground source heat pumps, and buried hot fluid pipelines and the surrounding soil. The TD of soil is governed by multiple factors, including density, water content, degree of saturation, organic content, and proportions of clay, sand, and silt, rendering its direct measurement both challenging and inherently uncertain. To address these limitations, the present study explores the application of advanced hybrid machine learning models integrating the extreme gradient boosting (XGBoost) algorithm for estimating TD in high-plastic soils. TD measurements were obtained using the dual-probe method (KD2-Pro) in a sunlight-free room under controlled temperature and humidity. A total of 180 experimental datasets comprising bentonite, bentonite–fly ash (silty sand) mixtures, and bentonite–sand mixtures were used to train and validate the models. The model performance and prediction accuracy were evaluated using several performance metrics, scatter plots, and regression error characteristic (REC) curves. The predictions confirm that the integration of metaheuristic optimization significantly enhances the performance of the baseline XGBoost model. Specifically, XGBoost-SSO outperforms the other models (training R² = 0.9744, testing R² = 0.9146), making it the most effective model for predicting soil thermal diffusivity. Moreover, SHapely Additive exPlanations (SHAP) analysis identified Sr (+0.072), Sand (+0.032), and Silt (+0.017) as the most influential features positively impacting the TD of high-plastic soil. While the developed models provide highly accurate forecasts, their inherent "black-box" nature presents interpretability challenges for engineering applications. To mitigate this, an open-source graphical user interface (GUI) was developed on the basis of the trained models, enabling practitioners to generate precise TD predictions efficiently.