Protocol for Clustering 4DSTEM Data for Phase Differentiation in Glasses

Mridul Kumar; Yevgeny Rakita

REVIEW 5 major objections 5 minor 18 references

Protocol for Clustering 4DSTEM Data for Phase Differentiation in Glasses

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Unsupervised clustering of 4D-STEM diffraction data separates a Ge-Sb-Te glass into four compositionally distinct regions without phase labels.

desk verdict A clear, honest demonstration that standard PCA+UMAP+k-means can produce four spatially coherent clusters in one GST 4D-STEM dataset, but the physical claim is only weakly supported and the clustering is done in UMAP space rather than in the original diffraction feature space. read the letter →

arxiv 2509.00943 v1 pith:YMDNYNCY submitted 2025-08-31 cond-mat.mtrl-sci cs.CVcs.LG

Mridul Kumar , Yevgeny Rakita This is my paper · ORCID

classification cond-mat.mtrl-scics.CVcs.LG MSC 62H3062P35 PACS 68.37.Ma07.05.Kf

keywords 4D-STEMphase-changematerialsGe-Sb-TeunsupervisedclusteringUMAPk-meanscompositionalheterogeneitydiffractionpatternanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a standard unsupervised machine-learning pipeline—preprocessing, PCA, UMAP/t-SNE embedding, then k-means with silhouette scoring—can sort the diffraction patterns collected across a Ge-Sb-Te phase-change glass into four spatially coherent groups. Because the diffraction patterns are unlabeled, the groups are not predefined phases; instead, the algorithm groups patterns by similarity, and the authors then show that the groups line up with distinct elemental signatures in simultaneously acquired EDS maps. Cluster 1 is oxygen- and germanium-rich, Cluster 2 is tellurium-rich, Cluster 3 is antimony-rich, and Cluster 4 is germanium-rich again, and averaged diffraction patterns differ across clusters. If correct, this gives materials scientists a practical, label-free prescription for mapping nanoscale compositional and structural heterogeneity in glasses and other complex materials from 4D-STEM data alone.

What carries the argument

The load-bearing object is the 4D-STEM dataset: a full electron diffraction pattern recorded at every scan position, here a 110×45 grid of 256×256 patterns (cropped to 70×45 after removing top and bottom artifacts). The argument runs through the vectorized diffraction pattern as the unit of similarity: patterns that resemble each other in high-dimensional feature space are taken to come from similar phases or compositions. UMAP is the mechanism that turns the PCA-reduced patterns into a low-dimensional geometry in which local and global neighbourhood relations are preserved, and k-means is the partition mechanism that cuts that geometry into discrete clusters. The silhouette score supplies t

What would settle it

Run the exact preprocessing and clustering pipeline on a homogeneous, compositionally uniform Ge-Sb-Te film of the same thickness. If the pipeline still returns four well-separated, spatially coherent clusters with the same chemical enrichments, then the clusters are artifacts of the measurement rather than phases. Alternatively, compare each cluster's averaged diffraction pattern with reference or simulated diffraction patterns for known GST phases and Ge oxide; if distinct clusters match distinct reference phases, the interpretation is confirmed.

Watch

Extended reading notes

Core claim

The central discovery claimed is that unsupervised clustering of 4D-STEM data can recover meaningful phase-like differentiation in an unlabelled Ge-Sb-Te glass. Starting from six scans of the same sample, the authors crop artifacts, remove the central beam, calibrate and center each 256×256 diffraction pattern, then reduce the 65,536-dimensional feature space to 90 principal components that retain over 95% of the variance. UMAP, chosen over t-SNE for better global-structure preservation and scalability, embeds these features in three dimensions, and k-means clustering on that embedding—with k=4 selected by the best average silhouette score—assigns every scan position to one of four clusters.

Load-bearing premise

The physical interpretation depends on the assumption that groups of similar diffraction patterns correspond to similar phases or compositions, rather than to thickness variations, scan drift, beam damage, or oxidation.

Editorial extensions

If this is right

The same PCA–UMAP–k-means recipe can be applied to other phase-change materials and glassy samples where no phase labels exist, turning raw 4D-STEM data directly into spatial phase maps.
For Ge-Sb-Te, the four clusters give concrete regions to correlate with EDS and with structural fingerprints, so local chemistry and structure can be studied at the ~2.5 nm probe scale without prior phase knowledge.
The result implies that the diffraction signal itself carries enough compositionally sensitive information to separate regions such as tellurium-rich and antimony-rich zones, so EDS may not be required for initial heterogeneity screening.
The choice of k=4 sets a testable prediction: this sample contains four distinct environments; subsequent high-resolution structural analysis can check whether these correspond to known GST phases, Ge-oxide regions, or intergranular zones.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The authors analyse one representative file in detail; a natural extension not reported is to run the pipeline on all six scans and check that the same four clusters appear in consistent spatial positions, which would distinguish intrinsic phases from scan-specific artifacts.
Because the EDS maps are used only after clustering to interpret the clusters, the direction of the claim is correlation, not prediction; the stronger claim that diffraction alone can identify composition would need a hold-out test where clusters trained on one region map a fresh region.
If the protocol is validated further, it could be paired with in-situ heating or electrical-switching experiments, where cluster maps at successive temperatures would track amorphous-to-crystalline phase separation as the material switches.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

The real content here is modest but genuine: the authors take a standard 4D-STEM workflow — PCA down to 90 components, UMAP to 3D, k-means, silhouette scoring — and show that on one GST scan it yields four clusters that map to spatially contiguous regions with visibly different EDS histograms and averaged diffraction patterns. That is a reasonable existence proof, not a new method or a new phenomenon. The paper is clearly written, the preprocessing is described concretely, and the authors are candid that t-SNE was tried first and abandoned for sensible reasons. The self-citation (ref 17) is not a problem.

The soft spots are real and central. The clustering is done on the UMAP embedding, and the silhouette score that picks k=4 is computed in that embedding. UMAP can create apparent clusters in continuous data, so the four-cluster structure may be an artifact of the embedding rather than of the diffraction data. The stress-test note is on target. The fix is straightforward: run k-means on the PCA scores directly, compare cluster assignments, and report UMAP stability across hyperparameters and seeds. Without that, the averaged diffraction patterns and chemical histograms in Figs. 3–4 don't yet prove distinct physical phases.

There are also validation gaps that are more conventional. No ground-truth phase labels, no error bars on the EDS histograms, silhouette score around 0.4, and the oxygen-rich Cluster 1 is just as consistent with surface oxidation as with an intrinsic phase. The analysis appears to be a single file (0034), even though six scans were taken; showing that the same clusters appear in other files would materially strengthen the claim. No data or code is released, which is a particular weakness for a paper whose stated contribution is a protocol.

The citation pattern is fine, and the authors are honest about the unsupervised framing. I disagree with the abstract's confidence in "four distinct clusters" as physical truth — that is an interpretation, not a demonstrated result. But the workflow itself is reproducible in principle and the empirical observation is worth taking seriously.

Who is this for? Microscopists working with 4D-STEM who want a starting recipe for phase segmentation, not readers looking for fundamental advances. I would send it to peer review, but with a clear request for major revisions: validate clustering on PCA space, test UMAP stability, show results across more than one scan, add statistical error bars, and release code/data. If those revisions land, the paper would be a useful methods reference. As it stands, it is a promising demonstration that overclaims its own certainty.

My advice: do not desk-reject it; send it for review with a strong request for the validation experiments above.

Referee Report

5 major / 5 minor

Summary. The manuscript presents an unsupervised machine-learning workflow for analyzing 4D-STEM diffraction data from a Ge-Sb-Te phase-change material. After cropping, central-beam removal, and PCA (90 components, ~95% variance), t-SNE and UMAP are used for visualization, and k-means is applied to the 3-component UMAP embedding. Silhouette scoring selects k=4, and the resulting cluster labels are mapped back to real space and compared with EDS maps. The authors report cluster-specific elemental signatures (O+Ge in Cluster 1, Te in Cluster 2, Sb in Cluster 3, Ge in Cluster 4) and claim that averaged diffraction patterns confirm structural differences. The paper is framed as a generalizable protocol for phase differentiation in glasses and disordered materials.

Significance. If validated, the protocol would be a useful demonstration of how unsupervised 4D-STEM analysis can reveal nanoscale compositional heterogeneity without phase labels, and it uses an accessible open-source software stack. The authors are careful to treat the problem as unsupervised and to validate clusters by mapping them back to real space and checking against independent EDS maps. However, the central physical claim is not yet supported: the clustering is performed and evaluated entirely within a nonlinear UMAP embedding, and the chemical and structural interpretations rest on visual comparisons of Gaussian-fitted EDS histograms and average diffraction patterns without quantitative uncertainties or significance tests. The reported silhouette score (~0.4) is moderate, and no evidence is provided that the four clusters are stable across UMAP hyperparameters or that they exist in the original PCA feature space. With additional validation, this could become a solid methodological contribution.

major comments (5)

[Application of k-Means Clustering; Figure 5] The k-means algorithm is applied to the 3-component UMAP embedding, and the silhouette score used to select k=4 is computed in that embedding (Figure 5). UMAP is a nonlinear visualization technique known to produce apparent clusters in continuous or noisy data; silhouette computed on the embedding does not validate clusters in the original diffraction feature space. The authors should report silhouette scores computed on the PCA scores (or on the original diffraction features), compare k-means results directly on PCA scores, and demonstrate stability across UMAP random seeds and hyperparameters (n_neighbors, min_dist). Without this, the claim of 'four distinct clusters' may be an artifact of the embedding.
[Results and Discussion; Figure 4A/B] The elemental intensity histograms are Gaussian-fitted and the peak values are plotted as atomic fractions, but no error bars, confidence intervals, or significance tests are provided. The reported differences (e.g., O and Ge enrichment in Cluster 1) may be within histogram width or may reflect surface oxidation rather than a distinct phase. The authors should provide quantitative statistics (mean ± std, p-values, or effect sizes) for the EDS distributions per cluster, and ideally test whether the O-enriched Cluster 1 spatially coincides with the oxidized regions cropped during preprocessing.
[Results and Discussion; Figure 4C] The averaged diffraction patterns are claimed to 'confirm structural variations,' but the support is only visual inspection. This is a load-bearing part of the central claim. Please provide quantitative comparisons, for example radial integration, peak positions/intensities, or comparison with simulated or reference patterns for amorphous/crystalline Ge-Sb-Te phases. Without such measures, the structural interpretation is not established.
[Materials and Methods; Preprocessing, Code 1] There is an inconsistency in dataset sizes: the text states that the final cropped dataset is (70,45,256,256) = 3150 diffraction patterns, but Code 1 reshapes X from (4950,256,256) to (4950,65536). Moreover, the detailed analysis appears to use only file 0034, while the abstract and introduction refer to six scans (0033-0038). If the protocol is meant to be reproducible across multiple scans, the authors should show cluster assignments and chemical signatures for more than one file, or clearly state that only the 0034 file is used for the demonstration.
[Results and Discussion; 'diffraction patterns with similar characteristics...'] The physical interpretation assumes that similarity in diffraction features is dominated by compositional or phase differences, rather than by specimen thickness, scan drift, beam damage, or oxidation. This is an untested assumption and is central to the claim that the clusters represent phases. A concrete test would be to compare the cluster map with a thickness-sensitive signal (e.g., low-angle scattering or ADF intensity) and to examine whether the observed clusters are stable when the analysis is restricted to a single scan or to regions away from the oxidized edge.

minor comments (5)

[Table 1] The table lists 'R Pixel Size' and 'Q Pixel Size' without defining whether 'R' is real-space and 'Q' is reciprocal-space. Please clarify units and definitions.
[Figure 1 caption] The caption states 'Red : Ge, Green : Sb, Blue : Te' but the figure text also says 'Blue regions are rich in Te' and 'TSNE Applied... Ge rich region show up as red colour.' The color scheme is confusing; please make it consistent and explicit.
[Application of UMAP] The text says UMAP was 'configured with k nearest neighbours and a minimum distance parameter of 0.1' but the actual value of k/n_neighbors is not specified. Please give the exact hyperparameters and, if possible, the random seed for reproducibility.
[Materials and Methods / Code] Only a PCA code snippet is provided. For a protocol paper, releasing the full preprocessing and clustering pipeline (or a link to a repository) would substantially strengthen reproducibility. This is not required for the scientific claim but would improve the manuscript.
[Throughout] Several typos and figure-label errors: 'clutered' in Figure 2 caption, 'diffractio n' in Figure 5, 'k nearest neighbours' should be 'n_neighbors'. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: clusters are independently validated against EDS maps and mean diffraction patterns; the only self-citation is non-load-bearing.

full rationale

The paper's derivation chain is: 4D-STEM diffraction patterns → crop/center/calibrate → PCA (95% variance) → UMAP/t-SNE embedding → k-means clustering → back-projection to real space → comparison with EDS elemental histograms and mean diffraction patterns. The clustering is performed entirely on diffraction-derived features; the EDS maps are used only after clustering, as an external validation set. There is no fitted parameter that is renamed as a prediction: the chemical signatures (O/Ge in Cluster 1, Te in Cluster 2, Sb in Cluster 3, Ge in Cluster 4) are descriptive statistics of EDS intensities per cluster, not predictions from the clustering algorithm. The only self-citation is reference 17, used for Table 1 instrument parameters (spot size, pixel sizes, bit depth); this is not load-bearing for the scientific claim. The UMAP-silhouette concern raised by a skeptical reader is a legitimate methodological risk (k-means on a nonlinear embedding, silhouette computed in embedding space, no hyperparameter stability shown), but it is a correctness/validity issue, not circularity: the cluster labels are not defined in terms of the EDS maps or of the final interpretations. No equation in the paper equates an output to an input by construction, and no uniqueness theorem or ansatz is imported through self-citation. Therefore the central claim retains independent content and the paper is not circular beyond a minor, inconsequential self-citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

Central claim rests on two domain assumptions: diffraction similarity corresponds to phase or composition similarity, and EDS maps can be matched to diffraction clusters. Neither is externally verified; the 95% variance PCA choice and silhouette-based k = 4 are modeling decisions.

free parameters (4)

PCA n_components = 90
Set to the smallest number achieving at least 95% cumulative explained variance (Code 1); affects all downstream embeddings and clusters.
UMAP min_dist = 0.1
Chosen by hand in the UMAP configuration; controls local versus global structure in the embedding.
UMAP n_neighbors = not reported
Text says 'configured with k nearest neighbours' but the numerical value is not stated, leaving the embedding under-specified.
Number of clusters k = 4
Selected by silhouette score over k from 2 to 10; no independent phase label verifies this choice.

assumptions (4)

domain assumption Similarity of diffraction patterns implies similarity of phase or elemental composition
Stated in Results and Discussion first paragraph: 'diffraction patterns with similar characteristics, likely corresponding to similar phases or elemental compositions are expected to be grouped together.'
domain assumption The 95% explained variance PCA truncation preserves the information needed for physical clustering
Code 1 sets n_components by a 95% cumulative variance threshold; this is a modeling choice, not a derived guarantee.
domain assumption EDS maps can be spatially matched to 4D-STEM clusters
The paper overlays cluster labels and EDS maps without discussing drift or resolution differences between the two measurements.
domain assumption Silhouette score maximum identifies physically meaningful cluster number
k = 4 is chosen by silhouette; no independent phase labels verify the choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Protocol for Clustering 4DSTEM Data for Phase Differentiation in Glasses." pith.science (2026). https://pith.science/paper/YMDNYNCY

@misc{pith2026250900943,
  author       = {Pith},
  title        = {Pith review of: Protocol for Clustering 4DSTEM Data for Phase Differentiation in Glasses},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YMDNYNCY}},
  note         = {Machine review of arXiv:2509.00943}
}

read the original abstract

Phase-change materials (PCMs) such as Ge-Sb-Te alloys are widely used in non-volatile memory applications due to their rapid and reversible switching between amorphous and crystalline states. However, their functional properties are strongly governed by nanoscale variations in composition and structure, which are challenging to resolve using conventional techniques. Here, we apply unsupervised machine learning to 4-dimensional scanning transmission electron microscopy (4D-STEM) data to identify compositional and structural heterogeneity in Ge-Sb-Te. After preprocessing and dimensionality reduction with principal component analysis (PCA), cluster validation was performed with t-SNE and UMAP, followed by k-means clustering optimized through silhouette scoring. Four distinct clusters were identified which were mapped back to the diffraction data. Elemental intensity histograms revealed chemical signatures change across clusters, oxygen and germanium enrichment in Cluster 1, tellurium in Cluster 2, antimony in Cluster 3, and germanium again in Cluster 4. Furthermore, averaged diffraction patterns from these clusters confirmed structural variations. Together, these findings demonstrate that clustering analysis can provide a powerful framework for correlating local chemical and structural features in PCMs, offering deeper insights into their intrinsic heterogeneity.

Figures

Figures reproduced from arXiv: 2509.00943 by the authors.

**Figure 2.** This figure shows the application of UMAP on reduced data of 0034 file and the clustering that comes out of it. It is clear that the regions rich in Ge are now clustered much better compared to t-SNE. Here, A. Shows the 3D projection of the data and B. Shows 2D projection of the data. 0 10 20 30 40 50 60 0 10 20 30 40 0 10 20 30 40 50 60 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0034 File After Clustering and Creati… view at source ↗

**Figure 3.** This figure shows the segregation of the real space of the sample. A. HAADF of the sample with maxpooling for creating the virtual image. Projection of the clustering classes onto the sample after application of k-means clustering. B. It shows the EDS maps of the elements in file 0034. 3/8 [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: A. Application of Gaussian Fit on the intensity histograms of EDS values corresponding to different classes obtained from the k-means clustering. B. Maximum atomic fraction of all the elements in different clusters. C. Mean diffraction patterns from the obtained cluste…

**Figure 5.** Figure 5: This figure shows the various steps involved in training the machine learning model for the determination of different phases in the sample. Since the available phases of the sample are not labelled for specific regions, and only the 4D-STEM diffraction patterns and ED…

**Figure 6.** Figure 6: This figure shows the collection of diffraction patterns in feature space. Points which are closer to each other are closely related. Application of UMAP To address the issues with t-SNE, we subsequently applied Uniform Manifold Approximation and Projection (UMAP)11, 1…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

[1]

Jones, R. O. Phase change memory materials: Why are alloys of Ge, Sb, and Te the materials of choice? Solid State Sci. 152, 107524 (2024)

work page 2024
[2]

Thermophysical Properties and Measuring Technique of Ge-Sb-Te Alloys for Phase Change Memory (Springer, 2020)

Lan, R. Thermophysical Properties and Measuring Technique of Ge-Sb-Te Alloys for Phase Change Memory (Springer, 2020)

work page 2020
[3]

Ortega, E. O. et al. Material Characterization Techniques and Applications (Springer, 2022). 7/8

work page 2022
[4]

X-ray diffraction (XRD) techniques for materials characterization

Epp, J. X-ray diffraction (XRD) techniques for materials characterization. In Materials Characterization Using Nonde- structive Evaluation (NDE) Methods , 81–124 (Elsevier, 2016)

work page 2016
[5]

& Yang, Z

Tang, CY. & Yang, Z. Transmission electron microscopy (TEM). In Membrane Characterization, 145–159 (Elsevier, 2017)

work page 2017
[6]

Savitzky, B. H. et al. py4DSTEM: A software package for four-dimensional scanning transmission electron microscopy data analysis. Microsc. Microanal. 27, 712–743, DOI: 10.1017/S1431927621000477 (2021)

doi:10.1017/s1431927621000477 2021
[7]

Savitzky, B. H. et al. py4DSTEM: Open source software for 4D-STEM data analysis. Microsc. Microanal. 25, 124–125, DOI: 10.1017/S1431927619001351 (2019). https://academic.oup.com/mam/article-pdf/25/S2/124/48323359/mam0124. pdf

work page doi:10.1017/s1431927619001351 2019
[8]

Sadri, A. et al. Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy. NPJ Comput. Mater. 10, 243 (2024)

work page 2024

Show all 18 references

[9]

Mason, R. L. & Gunst, R. F. Selecting principal components in regression. Stat. & probability letters 3, 299–301 (1985)

1985
[10]

& Hinton, G

van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. J. machine learning research 9, 2579–2605 (2008)

2008
[11]

& Melville, J

McInnes, L., Healy, J. & Melville, J. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018). 1802.03426

2018 arXiv
[12]

Ostrovsky, R., Rabani, Y ., Schulman, L. J. & Swamy, C. The effectiveness of Lloyd-type methods for the k-means problem. J. ACM (JACM) 59, 1–22 (2013)

2013
[13]

Allen, F. I. et al. Fast grain mapping with sub-nanometer resolution using 4D-STEM with grain classification by principal component analysis and non-negative matrix factorization. Microsc. microanalysis 27, 794–803 (2021)

2021
[14]

Shi, C. et al. Uncovering material deformations via machine learning combined with four-dimensional scanning transmis- sion electron microscopy. npj Comput. Mater. 8, 114 (2022)

2022
[15]

& Johnson, I

Wattenberg, M., Viégas, F. & Johnson, I. How to use t-SNE effectively.Distill 1, e2 (2016)

2016
[16]

Becht, E. et al. Dimensionality reduction for visualizing single-cell data using UMAP. Nat. biotechnology 37, 38–44 (2019)

2019
[17]

& Rakita, Y

Kumar, M. & Rakita, Y . Prediction of EDS maps from 4DSTEM diffraction patterns using convolutional neural networks (2025). 2508.20657

2025 arXiv
[18]

Wei, J. et al. Machine learning in materials science. InfoMat 1, 338–358 (2019). 8/8

2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Jones, R. O. Phase change memory materials: Why are alloys of Ge, Sb, and Te the materials of choice? Solid State Sci. 152, 107524 (2024)

work page 2024

[2] [2]

Thermophysical Properties and Measuring Technique of Ge-Sb-Te Alloys for Phase Change Memory (Springer, 2020)

Lan, R. Thermophysical Properties and Measuring Technique of Ge-Sb-Te Alloys for Phase Change Memory (Springer, 2020)

work page 2020

[3] [3]

Ortega, E. O. et al. Material Characterization Techniques and Applications (Springer, 2022). 7/8

work page 2022

[4] [4]

X-ray diffraction (XRD) techniques for materials characterization

Epp, J. X-ray diffraction (XRD) techniques for materials characterization. In Materials Characterization Using Nonde- structive Evaluation (NDE) Methods , 81–124 (Elsevier, 2016)

work page 2016

[5] [5]

& Yang, Z

Tang, CY. & Yang, Z. Transmission electron microscopy (TEM). In Membrane Characterization, 145–159 (Elsevier, 2017)

work page 2017

[6] [6]

Savitzky, B. H. et al. py4DSTEM: A software package for four-dimensional scanning transmission electron microscopy data analysis. Microsc. Microanal. 27, 712–743, DOI: 10.1017/S1431927621000477 (2021)

doi:10.1017/s1431927621000477 2021

[7] [7]

Savitzky, B. H. et al. py4DSTEM: Open source software for 4D-STEM data analysis. Microsc. Microanal. 25, 124–125, DOI: 10.1017/S1431927619001351 (2019). https://academic.oup.com/mam/article-pdf/25/S2/124/48323359/mam0124. pdf

work page doi:10.1017/s1431927619001351 2019

[8] [8]

Sadri, A. et al. Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy. NPJ Comput. Mater. 10, 243 (2024)

work page 2024