EUVIMEDEUVIMED entdecken
Crossref

PtyRANNOSAUR: Ptychography using Robust Artificial Neural Networks Optimized for Sub-Ångstrom Accuracy and Ultrafast Reconstruction

Rahim Raja, Kieran Loehr, Xiaochuan Ding, Jeffrey Huang, Bryan Clark, Pinshane Y Huang

Microscopy and Microanalysis · 27.07.2026 · 32 (Supplement_1)

DOI: 10.1093/mam/ozag053.984 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Microscopy and Microanalysis, online erschienen am 27.07.2026. Verfasst von Rahim Raja und 5 weiteren Autor:innen. Die Arbeit nennt 15 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Electron ptychography has emerged as a versatile phase-retrieval imaging technique capable of exceeding Scanning Transmission Electron Microscopy (STEM) resolution capabilities whilst offering improved dose efficiency – a capability achieved even in uncorrected electron microscopes [1,2]. Electron ptychography reconstructs a specimen's complex transmission function from a four-dimensional STEM (4D-STEM) dataset in which convergent beam electron diffraction patterns are recorded at each probe position. By exploiting the redundancy introduced through overlapping diffraction disks and dense probe sampling, ptychography enables recovery of phase information, which is otherwise lost when diffraction intensities are measured. Despite these advantages, ptychography's widespread adoption has been limited by its computational cost. Conventional approaches rely on iterative, physics-based solvers that repeatedly simulate electron wave propagation and update the specimen estimate to match experimental measurements [3]. These reconstructions typically require minutes to hours per dataset and may require extensive manual tuning of algorithmic hyperparameters to ensure sufficient convergence. This limits experimental throughput and precludes real-time feedback during data acquisition. Consequently, only a small fraction of collected data can typically be reconstructed, restricting analysis to limited fields of view. These challenges become more pronounced for large-area scans and for thicker or structurally complex specimens. Recent advancements in machine-learning (ML) abilities offer a solution to the temporal limitations by accelerating or replacing iterative reconstruction algorithms [4-10]. Data-driven models, such as deep neural networks, have demonstrated the ability to learn efficient mappings between raw measurements and reconstructed phase objects. The key advantage of these approaches is that they require single or few passes to retrieve a satisfactory reconstruction, which takes seconds to minutes. Integrating ML with electron ptychography reliably, however, presents the challenge of needing to respect physical constraints, generalize to experimental conditions, and consistently maintain fidelity to the materials structure, especially for thicker specimens. We introduce an ML-based approach to electron ptychography, combining the wave-optical modeling from the multislice method [11] with the efficiency of modern ML architectures. As we outline in Fig 1, our model relies on simulated 4D-STEM datasets generated using the multislice formalism [12] on modified crystal structures from the Materials Project database [13]. We use realistic variations in probe parameters, dose levels, and specimen structure to train a neural network to implicitly learn the relationship between simulated diffraction measurements and simulated real-space phase objects. Our approach produces rapid and robust reconstructions while remaining resilient to experimentally relevant variations in acquisition conditions. In Fig 2, we show examples of experimental datasets reconstructed using PtyRANNOSAUR and using ptyRAD [14], an iterative ptychography package. For both thin 2D materials, and thicker specimens, PtyRANNOSAUR provides comparable resolutions of up to 0.4 Å, within 30 seconds, including probe position correction capabilities. An iterative approach on these datasets can require up to 10 minutes for a rough reconstruction; however, it can require hours for probe position correction to stabilize and produce the highest quality reconstructions. Beyond computational speedup, data-forward approaches to ptychography introduce the opportunity of frontloading models with more realistic physics to reframe the inverse problem to produce more feasible phase object reconstructions. Overall, this work establishes a foundation for physically grounded, high-throughput ML-based ptychography that accommodates increasingly realistic forward models and complex specimen behavior. By enabling rapid, high-resolution reconstructions, this approach bridges the gap between the scalability of conventional STEM imaging and the rich information retrieval capabilities of iterative electron ptychography to allow real-time data analysis and feedback capabilities during ptychography acquisition. These advances are essential to fully leverage next-generation STEM instrumentation with rapidly increasing data acquisition rates, opening new opportunities for near-live large-area electron ptychography for high resolution materials characterization and discovery [15]. Schematic for training and ptychographic reconstructions of local patches of objects using PtyRANNOSAUR. During training, an electron scattering forward process, the multislice algorithm, is used to create diffraction patterns from a given crystal structure at a specified probe position. A 5x5 grid (gold dots) of 25 diffraction patterns (red borders) are stacked together to form the input to the CNN which outputs a 2x2 Å patch of the projected potential (blue border) centered on the central probe position. During inference, the CNN is given diffraction patterns from experimental 4D-STEM data, and the output patch is the ptychographic reconstruction, which is then stitched to form the full object using cross-correlation based stitching. PtyRANNOSAUR and PtyRAD ptychographic reconstructions evaluated on experimental 4D-STEM datasets. PtyRANNOSAUR reconstructions take approximately 30 seconds to reconstruct. a) Gold nanoparticle on an amorphous structure. b) PtyRAD evaluated on the gold nanoparticle results in lattice being resolved more clearly in the gold nanoparticle. However, artifacts are present elsewhere in the image. c) Twisted WSe2 bilayer evaluated using PtyRANNOSAUR reconstructs with similar quality and structure to d) PtyRAD reconstruction. Fourier transforms of WSe2 using e) the neural network model and f) the iterative ptychography reconstructions demonstrate PtyRANNOSAUR’s ability to match information transfer abilities of iterative ptychography. Scale bars: 10 Å.

Zitieren

APA

Raja, R., Loehr, K., Ding, X., Huang, J., Clark, B., & Huang, P. Y. (2026). PtyRANNOSAUR: Ptychography using Robust Artificial Neural Networks Optimized for Sub-Ångstrom Accuracy and Ultrafast Reconstruction. Microscopy and Microanalysis, 32(Supplement_1). https://doi.org/10.1093/mam/ozag053.984

Vancouver

Raja R, Loehr K, Ding X, Huang J, Clark B, Huang PY. PtyRANNOSAUR: Ptychography using Robust Artificial Neural Networks Optimized for Sub-Ångstrom Accuracy and Ultrafast Reconstruction. Microscopy and Microanalysis. 2026;32(Supplement_1). doi:10.1093/mam/ozag053.984

BibTeX
@article{Raja2026,
  author = {Raja, Rahim and Loehr, Kieran and Ding, Xiaochuan and Huang, Jeffrey and Clark, Bryan and Huang, Pinshane Y},
  title = {PtyRANNOSAUR: Ptychography using Robust Artificial Neural Networks Optimized for Sub-Ångstrom Accuracy and Ultrafast Reconstruction},
  journal = {Microscopy and Microanalysis},
  year = {2026},
  volume = {32},
  number = {Supplement_1},
  doi = {10.1093/mam/ozag053.984}
}
Kostenlos starten →