Predicting the Strength of Composites with Computer Vision Using Small Experimental Datasets

Composite materials offer versatile properties readily tailored to diverse applications, yet predicting their mechanical behavior remains challenging due to the complex interplay between morphology and performance. Here, we address these challenges by leveraging convolutional neural networks (CNNs) to analyze X-ray computed tomography (CT) images of cold-sintered polymer-ceramic composites for correlation with mechanical performance. Morphological features as inputs for traditional machine learning models yielded limited predictive accuracy. Application of transfer learning from pretrained CNNs improved prediction accuracy. The model was refined with Bayesian hyperparameter optimization and ensemble learning techniques to achieve R² values of up to 0.94 on unseen data. Additionally, we utilized the Z-stack nature of X-ray CT imaging -- where multiple 2D slices from a 3D dataset provide additional insights -- to apply a meta-learning approach for final predictions. This method enabled the prediction of intrinsic 3D properties from 2D slices that improved R² values to 0.95. This work demonstrates alternative approaches with machine learning using small datasets to uncover morphology-structure-property relationships in composites and highlights the potential of computer vision in property predictions for materials development.