Application of machine learning algorithms to model soil thermal diffusivity

Soil thermal diffusivity (k) is an important thermal property that significantly affects ground energy storage and heat transfer. Direct measurements of soil thermal diffusivity are challenging due to sensor limitations and variable soil physical properties, such as water content, bulk density, mineralogy and texture. Therefore, indirect estimations of k are commonly used. In this study, the abilities of six machine learning (ML) models, including k-nearest neighbours (KNN), multilayer perceptron (MLP), support vector regression (SVR), decision tree (DT), random forest (RF) and gradient boosting decision tree (GBDT), to estimate k values are evaluated based on a compiled database consisting of 999 samples. The ML model performances are evaluated by three model performance indicators (i.e., coefficient of determination-R2, mean absolute error-MAE and root mean square error-RMSE). The GDBT model best estimates soil thermal diffusivity, showing the highest fitness (R2 = 0.99). Model estimation accuracies are determined for varying numbers of available model inputs, and recommended minimum numbers of model inputs needed to accurately estimate thermal diffusivity are tabulated. This study demonstrates the ability of ML models to estimate values of soil thermal diffusivity, and it provides reference information for future thermo-related soil science and soil engineering applications.