Interactive Framework for Real-Time 4DSTEM Analysis and Reconstruction

Le résumé fourni par la source

GPU-accelerated algorithms have enabled a new class of 4DSTEM reconstruction and analysis methods [1,2]. Despite increased access to computational resources, these methods are largely applied offline due to additional infrastructure required for real-time interactivity. While real-time human feedback can enhance analysis and experimentation, enabling such feedback during data acquisition remains challenging due to microscope integration requirements, additional software development, and complex analysis pipelines [3]. We present a framework that integrates 4DSTEM algorithms with human interactivity while minimizing infrastructure constraints (Figure 1). The framework provides a unified, reusable interactive layer that embeds real-time feedback into existing analysis workflows through custom interactive widgets [4]. These widgets run on microscope workstations, HPC clusters, cloud platforms, and personal laptops. They can be used either as standalone Python components or embedded within a Python notebook environment, preserving full flexibility for customization alongside existing software packages and scripts. We demonstrate two real-time 4DSTEM applications operating directly on streaming data. A CUDA-accelerated single-sideband phase retrieval (SSB) method [5,6] is applied to a 256 × 256 scan with a 192 × 192 DECTRIS ARINA detector, collected at the nano@stanford. We show the role of human interactivity in refined aberration search and reconstruction. Using the resulting SSB reconstructions, we perform drift correction on image pairs acquired from orthogonal scan directions [7]. Drift vectors are visualized in real time, enabling immediate evaluation and correction of sample drift during data acquisition using a fully differentiable model incorporating translational, affine, and non-rigid distortions. By embedding human feedback directly into the analysis loop with GPU-accelerated methods, this framework enables a new class of 4DSTEM workflows in which experimental and analytical decisions are made at the microscope, including complex experiments such as multislice ptychography and tomography. All components are open source [8], enabling the community to extend these capabilities to additional 4DSTEM methods [9]. Schematic diagram illustrating the role of interactive widgets in bridging 4DSTEM algorithms, researchers, and microscopes for real-time feedback. Screenshot of an interactive alignment widget embedded in a Jupyter notebook. The first row contains two simulated orthogonal scan images of MoS2, where the second image is automatically aligned and rotated to match the first frame. The second row shows the merged image after additional rotation applied through user input.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.