Interpretable deep learning for atomicity classification of platinum nanoclusters in STEM images

Abstract Determining the number of constituent atoms in metallic nanoclusters (NCs) directly from imaging is key to understanding how atomicity governs their size-dependent properties. Scanning transmission electron microscopy (STEM), which captures real-space images of materials with tunable magnification down to the atomic scale, provides an invaluable means to probe such structures. However, despite these advantages, automated and accurate identification of NC atomicity remains challenging, requiring robust extraction of features such as projected shape and contrast distribution from imaging data. To address this challenge, we present a deep learning framework that classifies platinum NCs (Pt n ; n = 19, 30, 41, 55, 70) using high-resolution aberration-corrected STEM images. A convolutional neural network extracts structural features that are separable in UMAP (Uniform Manifold Approximation and Projection) space, with class-specific focus visualized using Grad-CAM (Gradient-weighted Class Activation Mapping). The model achieves high accuracy, even for mixed-atomicity samples ( n = 19, 41, 70) on a shared substrate. To address domain shift, we apply fine-tuning with high-confidence pseudo-labels, significantly recovering performance. A dual-channel model integrating Local Contrast Normalization (LCN) filtering achieves a coefficient of determination of R ² = 0.94 ± 0.03, outperforming size-based classification. This framework automates atomic-scale classification from STEM images and advances autonomous workflows via real-time analysis and machine learning based decisions.