DATA-DRIVEN PREDICTION OF MECHANICAL STRENGTH AND ELECTRICAL RESISTIVITY IN GRAPHITE-MODIFIED CEMENTITIOUS COMPOSITES

The present study proposed a data-driven machine learning approach for the prediction of mechanical strength and electrical resistivity of graphite-modified cementitious composites. The experimental data analyzed consisted of two independent sets: 373 observations of unconfined compressive strength (UCS) and 416 observations of electrical resistivity (ER), comprising graphitic-material properties, proportions of the mixtures, processing conditions and curing parameters. Six regression models – Linear Regression, Decision Tree, Random Forest, Gradient Boosting/ XGBoost, Support Vector Regression and K-Nearest Neighbors – were tested with cross-validation and an 80/20 train-test split. The performance of the models was evaluated based on R2, MAE, MSE, and RMSE. Random Forest gave the best fit to the model for the UCS prediction with an R² of 0.888, MAE of 3.75, MSE of 25.10 and RMSE of 5.01, while the best model for the ER prediction was Decision Tree, with an R² value of 0.991, MAE of 40.20, MSE of 1623.00, and RMSE of 67.58. The curing age, water-to-cement ratio, graphitic thickness, and ultrasonication strongly affected the UCS value, while the graphitic content, water-to-cement ratio, surface area, use of superplasticizer, and conditions related to the measurement had a more pronounced effect on the ER value. In general, it is shown that nonlinear machine-learning models are capable of capturing complex material-property relationships and are useful for optimizing the multifunctional and self-sensing cementitious composites.