Explainable Deep Learning for TEM-Based Classification of Carbon Nanoparticles: A ResNet50 Transfer Learning Approach

Precise and explainable nanoparticle categorization from Transmission Electron Microscopy (TEM) images is critical for nanotechnology material assessment and quality checking. Cutting-edge automated methods suffer from limited annotated datasets and high morphological heterogeneity, making it challenging to model generalization and robustness. In this paper, an efficient focused deep learning pipeline with a pretrained ResNet50 backbone and customized classification layers and a large class-balanced augmentation scheme with Albumentations is proposed. The enrichment is geometric and photometric transformation tailored to augment the sparse TEM dataset to 200 images per class for compensating for data sparsity and variation. The dataset is divided into training (70%), validation (15%), and test (15%) sets to maintain balanced testing. Model training uses AdamW optimization, learning rate scheduling, and early stopping to prevent overfitting. New to this study, Grad-CAM is used as a post-hoc method of explainability to generate visual model prediction explanations that increase interpretability and clinical confidence. On the balanced classes test set, the pipeline had a satisfactory accuracy of 98.33 ± 0.10% after training the model separately five times with random seeds of varying values and giving the mean standard deviation of the critical performance measurements over the runs for robustness and reproducibility, with average precision, recall, and F1-scores all above 0.98 for Diamante and Multi-Walled Carbon Nanotube (MWCNT) classes, with virtually no misclassifications confirmed through the confusion matrix. Originality of this work resides in synergy-based blending of transfer learning, a strong boosting approach tailored to small TEM nanoparticle images,and inclusion of Grad-CAM explainability, offering an interpretable, stable, and scalable platform for automatic nanoparticle analysis. Future research will look to expand interpretability with further explainability techniques, providing real-time inference, and comparing the method to other microscopy techniques to make it more relevant in nanomaterials science and industry.