Micrographic Evaluation of Carbon Nanotube Cement Composites via Machine Learning and Convolutional Neural Networks

Abstract Carbon nanotube (CNT)/cement composite is a promising material for structural health monitoring, where CNT dispersion and content critically affect the conductive network for self-sensing. Still, a simple method to assess low CNT levels is lacking, limiting their large-scale production. This study applied machine learning (ML) and convolutional neural networks (CNN) to predict these factors from optical microscopic images. A custom 2D-CNN directly learned spatial features from cropped images, reaching higher accuracy (0.993) and F1-score (0.953) in distinguishing dispersion methods (shear-mixed or sonicated) and classifying CNT contents (0–0.4 wt%) compared to baseline ML models. Support vector machine (SVM) and light gradient-boosting machine (LGBM) achieved decent accuracy (0.980) and F1-scores (0.928–0.938) using color histograms but less consistent performance to random data splits. These models can be integrated into concrete production, particularly batch verification and surface inspection, enabling scalable quality control of CNT/cement composites.