Segmented Detector Digitization and the Role of Denoising for Increasing Achievable Temporal Resolution in Phase Characterization

Phase contrast and retrieval methods have become increasingly important for characterization of materials in scanning transmission electron microscopy (STEM) in recent years. Since physical detectors only detect the intensity of the electron wave function, all phase information in the transmitted electron beam is lost. Phase contrast and retrieval methods aim to recover the phase information from the electron beam by recording the convergent beam electron diffraction (CBED) pattern in the bright field region, using pixelated or segmented detectors. Pixelated detectors capture fine details in the CBED pattern by capturing an image of the pattern at every scan position, creating a 4D dataset. This technique, known as 4D-STEM, has been shown extensively to produce high quality phase reconstructions through many phase retrieval methods including integrated center of mass (iCOM) and iterative ptychography [1-4]. However, 4D-STEM datasets tend to be massive (multiple GB), and hamper the practicable scan speed due to a high read-out overhead of the CBED pattern image at each scan position. As a result, the achievable temporal resolution with pixelated detectors is limited. Achieving high temporal resolution requires implementation of very fast scan coils, and use of a segmented detector with minimal read-out overhead. Still, all detectors have a finite response time for each electron detection event, and when the dwell time reaches or goes below this response time that causes streaking artefacts in the image [5]. These artefacts appear because signal for a single electron event is recorded in multiple subsequent pixels [6], but may be removed by live, in-hardware digitization of the detector signal [7]. This digitization method thresholds the gradient of the detector signal to pick out electron detection events and passes the digitized image back to the image recording software. This removes the analog noise present in the detector, giving a digital signal with a true-zero noise-floor, and ensure that each electron detected is a digital one and thus reducing the impact of detector surface inhomogeneity. In this work, we combine the use of ultra-fast scan coils [8] capable of achieving dwell times down to a few tens of nanoseconds, with digitization of the signal from a segmented annular all-field (SAAF) [9] scintillator-based detector. Working with digital signals opens the possibility of using denoising algorithms specifically designed to work with the Poisson statistics of electron detection [10]. The raw COM image in Fig. 1. shows how noisy the data is before denoising when acquired using beam current 1.9pA and dwell time 200 ns; the semi-convergence angle was 30 mrad. Even when summing 21 frames, no atomic columns are visible to the human eye in the raw COM image, whereas in the Poisson denoised image in Fig. 1. the atomic columns become visible. As a result of the denoising, the quality of the iCOM reconstructions is also improved, see Fig. 2. Here, we show how the use of this denoising framework helps boost the achievable temporal resolution for low-dose characterization of STO. Further, we compare this with the reconstructions achievable using iterative ptychography for data from this detector with only four segments and compare how the inherent denoising of the iterative ptychography algorithm compares with the explicit denoising for iCOM phase retrieval [11]. Comparison of the COM for a sum of 21 consecutive digitized frames from four segments on a segmented annular all-field detector, showing how noisy the raw signals are. The Poisson denoising helps make the atomic columns visible. iCOM reconstructions from the sum of 7, 14, and 21 frames. Note how Poisson denoising of the segment signals makes the atomic columns more easily visible than in the reconstructions from raw data.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.