A ResNet50 Transfer Learning and Grad-CAM-Based Framework for Explainable TEM Nanoparticle Classification

Precise and comprehensible nanoparticle classification from Transmission Electron Microscopy (TEM) images is crucial for material characterization progress as well as nanotechnology quality control. Yet, deep learning methods are limited by the size of sparsely annotated datasets as well as the high morphological variability of nanoparticles, thus the challenging task of model generalization and interpretability. This research presents an interpretable deep learning model consisting of a pretrained ResNet50 architecture, fine-tuned on custom classification layers and augmented by a class-balanced data augmentation strategy executed with Albumentations. The augmentation pipeline utilizes varied geometric and photometric transformations to increase the sparse TEM dataset to 200 images per class to counteract sparsity as well as increase variability. The data were split 70% training, 15% validation, and 15% test to offer a fair assessment. The models were optimized with AdamW optimization, learning rate adaptation, and early stopping to enhance stability and avoid overfitting. Grad-CAM was utilized as a post-hoc explanation technique for model interpretability to project class-specific activation areas. The suggested framework recorded a 95.00% test accuracy, with precision, recall, and F1-scores of about 0.95 for Diamante and Multi-Walled Carbon Nanotube (MWCNT) classes, verifying its consistency. The integration of transfer learning, specific augmentation, and open AI represents an open, scalable, and generalizable TEM-based nanoparticle classification without humans, opening the door to future multi-class and real-time nanomaterial imaging applications.