Machine learning-assisted characterization of the thermal conductivity of cement-based grouts for borehole heat exchangers

This research utilized machine learning (ML) techniques to forecast the thermal conductivity (TC) of cement-based grouts for borehole heat exchangers. Nine commonly used ML models were established and tested. Additionally, the accuracy of the ML models was contrasted with three conventional models. The results demonstrate that the back propagation neural network (BPNN) model emerges as the optimum prediction model with its highest accuracy (e.g., an R 2 of 0.991 on the test dataset). In addition, the BPNN model outperformed the three conventional models, while showing a notable increase of 29.3 % in R 2 compared with the optimum conventional model (i.e., Hashin-Shtrikam model). Finally, the SHapley Additive exPlanations analysis was conducted to comprehensively evaluate the importance of each input variable, and to analyse the individual relationships of the TC with input features. In conclusion, the proposed ML model proves an effective tool for forecasting the TC of grouts for borehole heat exchangers. This advancement facilitates the practical design and selection of grouts, ultimately improving the performance of ground source heat pumps. • ML models were built to predict the thermal conductivity of grouts for BHEs. • The BPNN model shows a high accuracy of R 2 of 0.991 for predicting the TC of grouts. • The BPNN outperformed the conventional models in predicting the TC of grouts. • SHAP analysis was conducted to build feature importance ranking. • Relationships between input features and TC were explained using SHAP plots.