Thermal conductivity of bentonite-sand/fly ash composites for underground thermal infrastructure using neural network and ensemble tree-based machine learning approaches

The thermal conductivity of bentonite-based mixtures significantly influences the performance and life-span of underground thermal energy systems and radioactive waste geological repository. The Machine Learning (ML) predictive models provide precise and accurate estimation of soil’s thermal conductivity over costly measurements. This investigation evaluates the predictive performance of multilayer perception structures, adaptive neuro-fuzzy architectures and regularized gradient boosting framework for evaluating heat transmission matrices for bentonite-sand (B-S) and bentonite-fly ash (B-F) mixes. The studied input (influencing) parameters included dry density, water content, particle compositions (sand, silt and clay), quartz content, and plasticity index. To perform ML, total 216 dataset of thermal conductivity of B-S and B-F mixtures were measured using KD-2 pro dual probe. The XGBoost model predicts thermal conductivity more accurately than ANN and ANFIS with overall R 2 = 0.996, MAE = 0.0113 and RMSE = 0.0246. Furthermore, the study involved a comparison of the proposed ML models with the four different empirical equations for thermal conductivity. It can also be noticed that ML techniques are more superior to the empirical equations for the present dataset.