Evaluating thermal conductivity of soil-rock mixtures in Qinghai-Tibet plateau based on theory models and machine learning methods

The thermal property of the scum layer (soil-rock mixtures) has dominant influence on the heat exchange efficiency between the lower rock layer and the upper environment in the open-pit mines of the cold regions. This paper presents a series of thermal conductivity tests (560 samples) on the scum particle to investigate the coupling effects of ice (moisture) content, temperature, and particle size distribution on the thermal properties. Previously reported models (47 empirical or theoretical models) were adopted to predicate the thermal conductivity of soil-rock mixtures in order to validate the evaluation ability of these models under the wide testing ranges. The comparison results indicate that the theoretical models, normalized model and linear/non-linear models all can not fully predict experimental results under the wide testing conditions. Three machine learning algorithms were used in the assessment presentation for the thermal properties of soil-rock mixtures. The performance of three machine learning algorithms were contrastively examined by using three important indexes (the coefficient of determination (R2), the root mean square error (RMSE) and the relative error (RE)). Based on the evaluation results, the performance ranking of three machine learning algorithms can be listed (GA-BP > SVR > RFR). This investigation is a beneficial attempt for the large data analysis to introduce the machine learning method into the determination of the thermal conductivity of soil-rock mixture under complex conditions.