Convolutional neural network-based visual classification of melt-fraction levels in phase change materials

Phase change materials (PCMs) are widely used in thermal energy storage (TES) systems due to their high latent heat capacity and ability to regulate heat during charging and discharging cycles. However, conventional techniques for evaluating melt-fraction levels are often time-consuming, labor-intensive, and require specialized instrumentation. Results from a convolutional neural network (CNN)-based approach for automated classification of melt-fraction levels in a conventional PCM (PureTemp 29™) is presented in this study. The data for the image processing algorithm are obtained from PCM melting experiments performed using an infrared (IR) camera. IR images were recorded during controlled thermocycling experiments. These IR images were preprocessed and categorized into 11 melt-fraction classes (0%–100%). A custom CNN architecture consisting of three convolution, batch-normalization, pooling stages followed by fully connected layers (1024–1024–256) and a softmax classifier was trained to map IR images to melt-fraction levels. The proposed model achieved 100% training accuracy and 96% test accuracy which demonstrated a strong generalization performance. Analysis of the confusion matrix showed that most classes were classified accurately, with misclassifications concentrated between visually similar melt-fraction states. These findings highlight the potential of deep learning-based image analysis as a fast, low-cost, and scalable tool for real-time monitoring and characterization of PCM melt-behavior in TES systems.