Machine learning based prediction model for thermal conductivity of concrete

Abstract

Thermal conductivity, k, is an important property of concrete, and it influences the design and energy-efficiency of many concrete-based structures. Due to the requirement of sophisticated test procedures, experimental measurement of k of concrete for every such structure is impractical, and therefore, a model for prediction of k is demanded. For this purpose, a data-driven machine learning (ML) model was developed in this study. The dataset for model training was developed from the published literature that contained missing data. For the imputation of missing data, fractional hot-deck imputation (FHDI) provided better performance than manual and naïve methods and is proposed for the imputation of similar datasets. Evaluation of nine ML algorithms, feature selection, and tuning of hyperparameters led to a final multilayer perceptron (MLP) model with the highest prediction accuracy. The model was developed using the Maxout activation function and three hidden layers, each containing 100 neurons. It performed reasonably well on the training, validation, and an independent dataset with a mean absolute error of 0.07, 0.14, and 0.10 W/m-K, respectively. The developed model was incorporated for a case study on a typical mass concrete mixture. The results indicated that a combination of quartz sand and siltstone could increase k and the rate of cooling and consequently can reduce the probability of thermal cracking in a mass concrete element. The developed model can provide more similar quantitative information that can aid in informed decision-making for the construction of critical structural elements. Besides, the robustness of the model can further be improved by a larger training dataset.