Learning continuous neural representations enables scalable and high-fidelity electron microscopy

Electron microscopy (EM) is indispensable for investigating nanoscale structures and dynamics across physical and life sciences. Conventional frame-based representations, however, store detector measurements as discrete pixel arrays, generating redundant and noise-sensitive datasets that grow rapidly with detector resolution and acquisition rate. Here, we introduce electron microscopy data representation (ENCODE), which learns a continuous neural representation of EM data and stores it as neural-network parameters rather than as independent pixel values. Across multiple EM modalities, ENCODE achieves substantial compression while preserving fidelity for downstream analyses. Its continuous formulation also enables arbitrary-index querying, intermediate-frame interpolation, and self-supervised denoising. Applied to low-dose in situ scanning transmission electron microscopy (STEM) observations of vanadium-doped tungsten disulfide (V-doped WS₂), ENCODE reconstructs transient atomic migration associated with vacancy filling and local lattice relaxation that are obscured in raw measurements. These results establish continuous neural representations as a scalable framework for EM data storage and analysis, supporting interpolation of smoothly evolving states and analysis of dynamics in noisy measurements. © 2026 The Author(s).