Predicting Partially Disordered Nanostructures: Graph-Based versus Image-Based Approaches

Nanomaterials often exhibit a balance between order and disorder, with the highest performance emerging in partially disordered regimes. Silver nanowire (AgNW) networks exemplify this behavior, achieving optimal transport properties within a “Goldilocks zone” between regularity and randomness. A key challenge is predicting how local structural motifs propagate into large-scale organization, as large-area nanoscale imaging is costly and incomplete.

Here, we compare two complementary approaches to reconstruct partially disordered networks: a graph-based method that learns local-to-global expansion rules, and an image-based machine learning model that completes partially observed SEM images. The comparison reveals how structural information encoded in graphs differs from that captured by image-based representations. This study provides one of the first direct evaluations of graph versus image frameworks for predicting partially disordered nanostructures and introduces spatially aware metrics for quantitative assessment.