Detecting Spatial Ordering of Nanoparticles with Geometric Deep Learning

Nanoparticle dispersion in heterogeneous catalysts plays a critical role in catalytic performance.We propose a robust and generalizable graph neural network (GNN) approach that combines the edge convolutional operator (EdgeConv) with a graph attention (GAT) layer to classify dispersion patterns in palladium on carbon (Pd/C) catalysts.Our method leverages GNNs to operate directly on particle location data extracted from scanning electron microscopy (SEM) images, thereby avoiding reliance on image features that may introduce bias.The proposed method offers an advantage over traditional image-based approaches, which risk overfitting to visual characteristics of the image that are unrelated to the spatial distribution of the nanoparticles.We validate the performance of our GNN architecture on multiple Pd/C samples with distinct carbon support types, achieving an accuracy of 89.84%, and demonstrate that our approach can reliably identify dispersion defects.The results highlight the potential of GNNs as a promising alternative for structure-based analysis and quality assessment of nanomaterialbased catalysts.