DINO4DSTEM: A self-supervised framework for structural discovery in 4D-STEM

Abstract

Nanodiffraction using 4D-STEM has become a key technique for quantitative nanoscale structural mapping in materials research, yet interpreting its high-dimensional datasets in structurally complex materials remains a major bottleneck. Existing analysis workflows typically rely on structural models, manual annotation, predefined classes, or sample-specific heuristics, limiting their ability to characterize heterogeneous complex materials. Here, we introduce DINO4DSTEM, a self-supervised machine learning framework that automatically discovers structurally meaningful states directly from raw diffraction data. Without structural models, manual labels, a predefined number of classes, or system-specific parameter tuning, the framework learns representations that organize diffraction patterns by their intrinsic structural similarities, transforming large collections of low-dose measurements into quantitative nanoscale structure maps. Across diverse datasets, DINO4DSTEM consistently identifies the dominant structural degrees of freedom, providing segmentation without human supervision. We applied the framework to the crystallization of indomethacin, a beam-sensitive, multidomain pharmaceutical system, revealing that crystallinity emerges from a partially ordered precursor and spans a continuous spectrum of structural order. The discovered nanoscale structural states are mapped to reveal the evolution of order across the specimen quantitatively. By replacing task-specific analysis with general self-supervised representation learning, DINO4DSTEM provides a broadly applicable framework for quantitative nanoscale structural mapping in complex materials, enabling the discovery of emergent structural organization in heterogeneous, beam-sensitive systems.