Machine-Learning-Based Parameterisation of Soil Thermal Conductivity for Shallow Geothermal and Ground Heat Exchanger Modelling

Thermal conductivity is a key input parameter in geotechnical and shallow geothermal engineering, directly influencing the design, efficiency, and long-term performance of ground heat exchangers, energy piles, and ground-source heat pump systems. Reliable parameterisation of this property in sandy soils remains challenging due to nonlinear interactions between water content, bulk density, and soil structure. This study develops a machine-learning-based workflow for robust parameterisation of thermal conductivity in quartz-rich sands using a large, internally consistent laboratory dataset comprising 1716 samples, including 1455 moist measurements used for modelling, obtained from nationwide site investigations. Air-dry specimens were identified as laboratory-induced drying states and excluded to restrict the analysis to hydro-mechanical conditions representative of typical shallow subsurface environments. Several regression algorithms representing different modelling strategies were evaluated within a unified and reproducible framework and benchmarked against selected classical empirical formulations. Model performance was assessed using standard accuracy metrics together with diagnostics describing the functional stability of predicted thermal-conductivity surfaces. The results reveal a systematic trade-off between predictive accuracy and functional consistency, indicating that models optimised for accuracy may produce functionally unstable and less suitable parameterisations for engineering applications. Accuracy-optimised models frequently produce locally irregular parameter fields, whereas more strongly regularised models yield smoother and physically more coherent response surfaces. The proposed workflow supports reliable thermal-property parameterisation for geotechnical design and shallow geothermal modelling.