A reliable model to predict thermal conductivity of unsaturated weathered granite soils

The aim of this study is to present a new empirical model to predict thermal conductivity, based on experimental databases that might be particularly applicable to unsaturated weathered granite soils. To establish the experimental databases, a number of thermal conductivity tests were conducted using a new probe system combined with a volumetric pressure extractor. Thereby, 162 data were collected and applied to the prediction model. The prediction model has two empirical coefficients that determine the shape of a graph, and it is not easy to determine these coefficients unless experiments are conducted. Thus, in this paper, we propose using an artificial neural network model to obtain these empirical coefficients without experiments, given only information on the soil properties. Moreover, to evaluate the applicability of the trained network model, it was tested for data sets that had not been introduced during the training stage. According to the verification results, the trained network model presents reasonable prediction results (R2 = 0.9046), even for the new testing data. In addition, the model traces the measured data curve with fairly good agreement, depending on a sample's porosity, regional characteristics, and degree of saturation.