EUVIMEDEUVIMED entdecken
Crossref

A Cluster-Based Filtering Technique for Denoising 4D-STEM Datasets

Adan Mireles, Yimo Han

Microscopy and Microanalysis · 24.07.2024 · 30 (Supplement_1)

DOI: 10.1093/mam/ozae044.927 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Microscopy and Microanalysis, online erschienen am 24.07.2024. Verfasst von Adan Mireles und 1 weiteren Autor:innen. Die Arbeit wurde bisher 3-mal zitiert (Crossref) und nennt 3 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The introduction of direct electron detectors has facilitated the rapid acquisition of 4D-STEM datasets (Figure 1(a)), benefiting nanobeam 4D-STEM applications such as strain mapping. Nonetheless, the sensitivity to Poisson noise poses a notable obstacle, particularly regarding the precision required for strain measurements. In nanobeam 4D-STEM, the accuracy of strain measurements is intrinsically linked to the precise determination of Bragg peak centers of mass [1]. Various denoising techniques exist. However, they often come with compromises, such as reduction of resolution or artifact introduction. To overcome these challenges, this study introduces a novel denoising technique adapted for nanobeam 4D-STEM data. This technique begins with an unsupervised learning approach to cluster diffraction patterns, enabling the discrimination of different materials and phases within the dataset, as illustrated in Figure 1(b) [2]. This clustering is essential for the subsequent two-step denoising process involving background subtraction and cluster-specific median filtering. In the first step, diffraction patterns within each cluster are averaged, Bragg peaks are identified and removed, the resulting background is reconstructed through interpolation (as detailed in Figure 1(c)), and the result is then subtracted from the corresponding cluster. The second step employs a cluster-specific median filter (magnified in Figure 1(d)), which traverses the 4D dataset with a user-defined kernel size, adjusting pixel values in the central diffraction pattern based on comparisons with same-position pixels in adjacent diffraction patterns that belong to the same cluster. Those pixels that exceed the median by a user-defined threshold (a multiple of the median) are then selectively replaced by the median value, as demonstrated for a single diffraction pattern in Figure 1(e). We applied the denoising method to WS2-WSe2 lateral heterojunction data, where the inherent 4.5% lattice mismatch between the materials induces strain in the material with the larger lattice constant [1]. The aim was to assess the accuracy of strain measurements on three versions of the same dataset: raw, without background, and filtered (denoised). Figure 2(a) displays the clustering results of the experimental dataset, identifying three distinct materials within the dataset: amorphous SiNx, WS2, and WSe2. A qualitative comparison of a WS2 diffraction pattern across raw, background-subtracted, and denoised stages reveals significant improvements in detail visibility. Furthermore, the precision of center-of-mass calculations for 2nd-order Bragg peaks, critical for strain measurements, improved by 50% following denoising, as shown in the marginal distributions of the scatter plot in Figure 2(c). This figure shows the maximum error at each processing step using a convex hull, whose area diminishes with each subsequent step. This enhancement in precision underscores the denoising method's efficacy, with Figure 2(d) highlighting the improved discernment and separation of WS2 and WSe2 peaks within the data. In conclusion, we develop a cluster-selective median filter approach to enhance the precision of structural measurements such as strain. This capacity to retain high-resolution details facilitates deeper insights into material properties. This advancement marks a pivotal step in the characterization and comprehension of complex material systems, promising to enrich our understanding of material behavior at the nanoscale [3]. (a) 4D-STEM diffraction pattern captured using EMPAD. (b) Raw dataset and its background-free version showing two materials (outlined in yellow dash-dot and orange), with a 3×3 kernel (green) modifying the central pattern's pixels (red, dotted outline). (c) Background extraction: averaging within clusters, Bragg peak removal, and post-interpolation result. (d) Close-up of filter: adjustments made using same-cluster patterns. (e) Comparison of diffraction patterns before and after filtering. (a) Clustering results for the WS2-WSe2 lateral heterojunction data (scale bar 250 nm). (b) Comparison of a sample WS2 diffraction pattern in its raw, background-subtracted, and denoised forms. (c) Scatter plot and marginal distributions of centers of mass (CoM) for all 2nd-order WS2 Bragg peaks across the three dataset variations: raw (in red), background-subtracted (in green), and denoised (in blue). Scatter plot outlines correspond to each dataset’s convex hulls. (d) Strain histograms derived from the raw, background-subtracted, and denoised datasets.

Zitieren

APA

Mireles, A., & Han, Y. (2024). A Cluster-Based Filtering Technique for Denoising 4D-STEM Datasets. Microscopy and Microanalysis, 30(Supplement_1). https://doi.org/10.1093/mam/ozae044.927

Vancouver

Mireles A, Han Y. A Cluster-Based Filtering Technique for Denoising 4D-STEM Datasets. Microscopy and Microanalysis. 2024;30(Supplement_1). doi:10.1093/mam/ozae044.927

BibTeX
@article{Mireles2024,
  author = {Mireles, Adan and Han, Yimo},
  title = {A Cluster-Based Filtering Technique for Denoising 4D-STEM Datasets},
  journal = {Microscopy and Microanalysis},
  year = {2024},
  volume = {30},
  number = {Supplement_1},
  doi = {10.1093/mam/ozae044.927}
}
Kostenlos starten →