AI-guided Recognition of Local Patterns in Nanoparticles from Tomographic Images

This study presents an atom-centred extension of the Bayesian deep-learning framework ARISE for the automated structural analysis of nanoparticles reconstructed by atomic electron tomography (AET). By replacing the original box-based strided sampling scheme with spherical, atom-centred local environments encoded via SOAP descriptors, the workflow enables physically interpretable, uncertainty-aware classification of 108 crystal prototypes without retraining the pre-trained model. In addition to prototype probabilities and Bayesian uncertainty estimates, latent representations from the neural network are analysed using manifold learning (UMAP) and density-based clustering (HDBSCAN), linking learned structural features to real-space atomic configurations. The approach is benchmarked on a synthetic multi-grain polycrystal and subsequently applied to experimental Pd and Pd@Pt nanoparticles, revealing defect-related distortions and medium-range structural motifs at chemically diffuse interfaces. Overall, the study demonstrates that combining atom-centred descriptors, Bayesian uncertainty quantification, and unsupervised latent-space analysis provides a flexible and data-efficient framework for the exploratory structural characterization of complex nanomaterials.