Enhancing Machine Learning-Based Prediction of Electrical Conductivity in Cementitious Composites

Abstract In this study, machine learning approaches were used to predict the electrical conductivity (EC) of conductive cementitious composites under data-limited conditions. Data scarcity was addressed via data augmentation based on a tabular variational autoencoder (TVAE), and multiple machine learning models were trained and evaluated using a consistent set of input variables and validation procedures. Results showed that prediction performance differed across model types. Tree-based and ensemble learning models exhibited stable predictive behavior after data augmentation and hyperparameter optimization, whereas distance-, kernel-, and neural network-based models were more sensitive to changes in data distribution and optimization strategies. TVAE-based data augmentation had a model-dependent effect on prediction. In addition, Shapley additive explanations analysis indicated that the mixing amount of conductive materials and fly ash content were among the most influential factors governing EC, with fly ash content exhibiting nonlinear, value-dependent contribution patterns. These findings provide insights into model-dependent prediction behavior and variable contribution characteristics in EC prediction for conductive cementitious composites under data-limited conditions.