Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Fast 4D-STEM-based phase mapping for amorphous and mixed materials

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Randomized NMF makes phase mapping of amorphous 4D-STEM data up to 100x faster.

desk verdict A solid methods paper: randomized NMF on raw 4D-STEM data gives a real, measured speedup, but the 'minimal error' claim needs spectral evidence on real data. read the letter →

arxiv 2507.17068 v1 pith:N57LVJPN submitted 2025-07-22 cond-mat.mtrl-sci physics.ins-det

Andreas Werbrouck , Nikhila C. Paranamana , Xiaoqing He , Matthias J. Young This is my paper
classification cond-mat.mtrl-sciphysics.ins-det
keywords 4D-STEMphasemappingamorphousmaterialsrandomizednonnegativematrixfactorizationQBdecompositiontransmissionelectronmicroscopybatteryinterfaceTiO2
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 prepending nonnegative matrix factorization (NMF) with a randomized low-rank QB decomposition makes phase mapping of 4D-STEM data from amorphous and mixed materials practical. Classical NMF on raw diffraction data scales as O(nmk) with the number of probed positions, diffraction features, and components, which puts large datasets out of reach. With QB preprocessing, the scaling becomes independent of the largest data dimension, yielding speedups of up to two orders of magnitude while producing the same components as full NMF. Demonstrations on a TiO2 coating on SiO2 and a crystalline-amorphous battery interface show that the fast route can separate and map components that are invisible to standard virtual-detector analysis.

What carries the argument

The load-bearing object is the QB decomposition used as a preprocessing step for NMF: a random projection sketches the m×n data matrix V into a much smaller m×(k+q) matrix B, with q an oversampling parameter, after which B is factorized and the result is rescaled back through the orthogonal factor Q. This reduces the effective dimension entering the non-convex NMF optimization, changing the scaling from O(nmk) to roughly O(nk) and shifting most of the computational work into the randomized sketch. In the real-data examples, oversampling q=30 and four subspace iterations are used to keep the approximation error small.

What would settle it

Compute the singular-value spectrum of the real 4D-STEM matrices used here, or of a similar dataset with known amorphous phases, and check how much energy lies beyond the top k+q singular values; if the normalized QB reconstruction error is large, or if a known phase only appears when the oversampling or subspace iterations are increased, the claim that RNMF preserves the NMF solution for these data fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that randomized NMF—standard NMF applied after a QB decomposition of the data matrix—reproduces the factorization that full NMF would find, with reconstruction error effectively unchanged, while cutting computation time by one to two orders of magnitude on 4D-STEM datasets. Because the QB step compresses the largest dimension of the data matrix before factorization, runtime scales with the smaller dimension, making it feasible to factorize raw, unintegrated diffraction data rather than azimuthally integrated pair distribution functions. The paper demonstrates this on synthetic matrices, on an amorphous TiO2 film on SiO2 where a distinct low-density surface phase is resolved, and on a mixed crystalline NMC / amorphous LGPS battery interface where eight components map the interface region. The authors present the method as a first-pass virtual detector that works equally well for crystalline, amorphous, and mixed data.

Load-bearing premise

The method assumes the full 4D-STEM data matrix is well approximated by a low-rank subspace of dimension k+q; if the subtle diffraction differences that distinguish phases, such as strained interfaces or low-density surface layers, live mostly in modes beyond that sketch, the fast factorization can reproduce the overall data while losing exactly the components the analysis aims to map.

Editorial extensions

If this is right

  • A 4D-STEM dataset that previously took roughly 46 minutes to factorize into three components is reported to take 47 seconds with RNMF, a 60-fold speedup, enabling routine re-analysis with different component counts.
  • Phase mapping can be applied directly to raw, unintegrated diffraction data, preserving Bragg peak positions and orientation or strain information that azimuthal integration discards.
  • Amorphous and mixed materials, which lack sharp Bragg peaks, become accessible to the same NMF-based unmixing used for crystalline grain mapping.
  • Residual maps after factorization provide a practical, if imperfect, guide for choosing the number of components, since spatially structured residuals signal that more components are needed.
  • Because the bottleneck is lowered so dramatically, much larger 4D-STEM maps with more pixels become tractable to analyze on a laptop.

Reading between the lines

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

  • Beyond the paper: the same QB-preprocessed NMF should transfer to other count-based hyperspectral imaging modalities, such as EELS or EDS spectrum images, wherever the data are non-negative and approximately low-rank.
  • Beyond the paper: the paper's own caveat that the TiO2 surface phase might be a continuous depth-dependent density gradient rather than a discrete phase suggests a testable extension—compare RNMF components against a continuum model of varying density and check whether the two-component split is stable across oversampling settings.
  • Beyond the paper: combining RNMF with online or streaming acquisition could move phase mapping from post-processing to near-real-time feedback during STEM scans, since the cost per factorization drops enough to run while data are still being collected.
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.

Referee Report

4 major / 7 minor

Summary. The manuscript proposes Randomized NMF (RNMF), in which a randomized QB decomposition is used as a preprocessing step before classical non-negative matrix factorization, and applies it to 4D-STEM data of amorphous and mixed materials. The authors claim that QB preprocessing makes the scaling independent of the largest data dimension (~O(nk)), leading to speedups of one to two orders of magnitude on synthetic data and a 60-fold speedup on a real TiO2/SiO2 dataset, while producing the same components as full NMF. They demonstrate the method on synthetic data, on an amorphous TiO2 coating on SiO2 spheres, and on an NMC-LGPS battery interface, and they discuss pitfalls including the difficulty of choosing the number of components and the risk of interpreting continuous strain/thickness gradients as discrete phases.

Significance. If the speed and equivalence claims hold, RNMF is a practically useful first-pass tool for phase mapping of raw 4D-STEM data without azimuthal integration, which would be a genuine advance for amorphous and mixed materials where NMF is otherwise prohibitively slow. The paper is honest about known limitations of NMF, provides measured runtimes, and makes a software package available on GitHub. However, the theoretical complexity claim is overstated, and the central claim that QB projection introduces 'minimal' error and returns 'the same components' is only rigorously demonstrated on exactly low-rank synthetic data, not on the real datasets where it matters. These issues are addressable with additional experiments and corrected analysis, so the manuscript is promising but not yet ready in its current form.

major comments (4)
  1. [Abstract and Section II-C] The asymptotic scaling claim is inaccurate. The abstract states that QB preprocessing achieves scaling independent of the largest data dimension (~O(nk)), and Section II-C states scaling of ~O(mk) instead of ~O(nmk). However, forming the randomized QB sketch requires multiplying the m x n matrix V by an n x (k+q) random matrix, and the power iteration (VV^T)^p V Omega costs O(mn(k+q)p) operations; the subsequent QR factorization adds O(m(k+q)^2). The total cost therefore remains at least linear in both m and n, and the NMF on the reduced m x (k+q) matrix costs O(mk(k+q)) per iteration. The empirical speedups may be real, but the stated asymptotic advantage is not. Please correct the complexity analysis and clearly separate the cost of the QB pass from the cost of the NMF on the reduced matrix.
  2. [Sections III-A, III-B, III-C and Figure S1] The central claim that RNMF induces 'minimal error' and returns 'the same components' as full NMF is not established for the real datasets. In Section III-A, the synthetic matrices are exactly rank k by construction (outer products of k components), so the QB projection is exactly lossless by design; this does not test behavior on the TiO2/SiO2 and NMC-LGPS data, which are not exactly low rank. No singular value spectrum of V, no relative QB projection error ||V - Q Q^T V||_F / ||V||_F, and no quantitative component-similarity metric between RNMF and standard NMF are reported for the real data; Figure S3 is only visual. If the subtle diffraction signatures interpreted as the TiO2 surface phase or the strained NMC interface modes are carried by singular components beyond rank k+q, the preprocessing would remove exactly the signal that the phase maps claim to show. Please add these diagnostics for both real datasets with the actual parameters (k=3 and k=8, q=30, p=4).
  3. [Sections II-C and III-B] RNMF is randomized and NMF is non-convex, but the real-data results are reported from a single run and no random seed is given. A different random projection, or a different NMF initialization after the same projection, can in principle converge to a different local optimum. To make the claim of a stable 60x speedup with identical output reproducible, please report run-to-run variability, for example five independent RNMF runs with different random projections, including relative differences in W and H and agreement of the phase maps.
  4. [Sections IV-C and V] The paper's own discussion in Section IV-C contains a significant caveat that undercuts the materials conclusion. The authors write that the TiO2 decomposition into two components 'could be viewed as a continuous shift between the two phases as a function of depth rather than two distinct phases,' and that for the NMC-LGPS interface 'an infinite number of components can be used to explain the differences in spectra.' Given this, the Conclusion's statement that RNMF reveals 'a low-density TiO2 surface phase of 10 nm thickness' is stronger than the evidence presented in the main text. The SI PDF analysis may support this, but the main text should either state the quantitative evidence with uncertainties or soften the claim.
minor comments (7)
  1. [Figure 4 caption] The caption lists the spatial heatmaps as '(b,d,e, respectively)', but panels d and f are the heatmaps; the panel letters should be '(b,d,f)'.
  2. [Section II-C] The text uses 'example 0' to refer to the synthetic data, while Sections III-A uses 'Example 0'; please use consistent labeling throughout.
  3. [Section III-A and Figure 3] Please state explicitly whether the reported RNMF runtimes include the QB decomposition time; the text in Section III-A suggests that they do, but this should be stated in the Methods or the figure caption.
  4. [Section III-A] The statement that 'the reconstruction error was the same for both NMF and RNMF' should be quantified with a numerical value (for example, the relative Frobenius norm) rather than relying only on a visual comparison in Figure S1.
  5. [Section III-B] The phrase 'resulting in the same components and mapping (Figure S3)' is only a visual comparison; please provide a quantitative similarity measure or refer to the new diagnostics requested in the major comments.
  6. [Section III-C] The text says 'we present these (manually) grouped components' in Figure 5; please define the grouping criteria more precisely, since the choice of grouping affects the interpretation of the interface components.
  7. [Data and code availability] A statement on data and code availability for the RNMF benchmark and the two real 4D-STEM datasets is missing; the ePDF GitHub link is helpful but does not clearly cover the RNMF experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the speedup claim is an empirical benchmark and the phase mapping results are outputs of the factorization, not fitted inputs renamed as predictions.

full rationale

The central claim is that QB-randomized NMF is one to two orders of magnitude faster than standard NMF with minimal induced error. This is established in Example 0 by direct benchmark on synthetic matrices: the paper constructs datasets as outer products of k component matrices (exactly rank k), runs both NMF and RNMF, and reports runtime and equal reconstruction error. No parameter is fitted to a subset of data and then presented as a prediction; the speedup is a measured comparison. On the real 4D-STEM datasets, RNMF is used to produce a factorization, and the resulting components and maps are interpreted as phases. This is analysis of the decomposition output, not a prediction that is equivalent by construction to an input. The low-density TiO2 surface phase and NMC-LGPS interface components are interpretations of RNMF components, and the paper explicitly acknowledges the ambiguity in Section IV-C, noting that the two TiO2 'phases' could be a continuous shift rather than discrete components. That is an honest limitation, not circularity. Self-citations appear (e.g., the NMC-LGPS dataset is from reference [20], and earlier ePDF work is cited for materials preparation), but these are provenance/background citations and are not load-bearing for the algorithmic speedup or for any derivation that reduces to them. The skeptic's concern that the real data may not be low-rank and that no spectral decay is reported is a validation and generalization gap, not circular reasoning: the paper does not claim to derive the low-rank property from the data, and its conclusions about real data rest on the assumption rather than on a definitional equivalence. Under the criteria requiring a quoted exhibit of a specific reduction of a claim to its own input, no circular step can be identified.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central speedup claim rests on standard randomized linear algebra (Halko et al. 2011) and a low-rank assumption on the 4D-STEM data; the physical interpretation rests on an assumption that NMF components correspond to linearly mixed phases, which the authors themselves qualify in the discussion. The component count k and the QB oversampling and subspace settings are hand-chosen per dataset. No new physical entities are introduced.

free parameters (3)
  • Number of components k = 3 for TiO2/SiO2, 8 for NMC-LGPS
    Selected post hoc from reconstruction error elbow and residual spatial structure; acknowledged as an open problem in Section IV-C.
  • Oversampling parameter q = 20 (synthetic), 30 (examples 1 and 2)
    Chosen to balance speed and numerical accuracy of the QB decomposition; no sensitivity analysis is given.
  • Subspace iterations = 2 (synthetic), 4 (examples 1 and 2)
    Chosen to increase numerical accuracy of the QB decomposition; no sensitivity analysis is given.
assumptions (5)
  • domain assumption The measured 4D-STEM matrix V is approximately low-rank, so a random projection onto k+q dimensions preserves the structure needed for NMF.
    Invoked in Methods II-C via the QB decomposition; no spectral decay analysis of the real datasets is provided.
  • domain assumption NMF components can be interpreted as linearly mixed physical phases (diffraction fingerprints and spatial maps).
    This is the basis for phase mapping throughout Section III; the authors themselves note in Section IV-C that continuous strain and thickness effects break this assumption.
  • standard math Randomized QB decomposition error bounds from Halko et al. apply and are small enough for these datasets.
    Relied on as unproved background; cited as [23], but no explicit bound or verification is given on real data.
  • ad hoc to paper Residual spatial structure is a reliable indicator that more components are needed.
    Used in Sections III-B and III-C to justify choosing 3 and 8 components; the holdout method was tried and rejected in Section II-C, so this heuristic is introduced for this analysis.
  • domain assumption Raw electron diffraction counts are nonnegative and linearly additive, so NMF on raw 2D data is valid.
    The authors argue this in Section IV-B; it is physically standard for electron counts but not directly verified for these datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast 4D-STEM-based phase mapping for amorphous and mixed materials." pith.science (2026). https://pith.science/paper/N57LVJPN

@misc{pith2026250717068,
  author       = {Pith},
  title        = {Pith review of: Fast 4D-STEM-based phase mapping for amorphous and mixed materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N57LVJPN}},
  note         = {Machine review of arXiv:2507.17068}
}
abstract

All materials are made from atoms arranged either in repeating (crystalline) or in random (amorphous) structures. Diffraction measurements probe average distances between atoms and/or planes of atoms. A transmission electron microscope in scanning mode (STEM) can collect spatially resolved 2-dimensional diffraction data, effectively creating a 4-dimensional (4D) hyperspectral dataset (4D-STEM). Interpretation strategies for such 4D data are well-developed for crystalline materials, because their diffraction spectra show intense peaks, allowing for effective phase and crystal orientation mapping at the nanoscale. Yet, because of the continuous nature of the diffraction data for amorphous and mixed materials, it is challenging to separate different amorphous contributions. Nonnegative matrix factorization (NMF) allows separation of 4D-STEM data into components with interpretable diffraction signatures and intensity maps, independent of the structure. However, NMF is a non-convex optimization problem and scales ~ O(nmk) with n the number of positions probed, m the number of diffraction features and k the number of components, making analysis of large 4D datasets inaccessible. Here, we apply QB decomposition as a preprocessing step for NMF (Randomized NMF or RNMF) to achieve scaling independent of the largest data dimension (~O(nk)), opening the door for NMF analysis of 4D-STEM data. We demonstrate our approach by mapping a thin TiO$_2$ layer on top of SiO$_2$, and a LiNi$_{0.6}$Co$_{0.2}$Mn$_{0.2}$O$_{2}$ (NMC) - Li$_{10}$GeP$_2$S$_{12}$ (LGPS) mixed crystalline-amorphous battery interface, illustrating strengths and limitations of using RNMF for structure-independent phase mapping in 4D-STEM experiments.

Figures

Figures reproduced from arXiv: 2507.17068 by the authors.

Figure 1
Figure 1. a) 4D-STEM working principle. b) Dataflow diagram of our approach [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Comparison of the execution time of RNMF and NMF on synthetic [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. RNMF decomposition of ALD TiO2 on a SiO2 sphere, including the diffraction signal, W, for components 1-3 (a,c,e respectively), and the corresponding spatial heatmap, H, of components 1-3 (b,d,e, respectively). Panels g and h show rreciprocal, and rreal, respectively. We note that the intensity for each diffraction fingerprint a, c, e and g is rescaled, such that the scalebars in b, d, f and h reflect the relative vo… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Phase maps of the interface between NMC and LGPS. Eight components were used to analyze this 4D-STEM dataset with RNMF. Components were [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Using superpixels for interpretable feature reduction in large 2D diffraction datasets

    cond-mat.mtrl-sci 2026-07 conditional novelty 6.0 of 10

    Variable-size superpixel averaging of 2D diffraction data reduces feature count and accelerates NMF phase mapping by over 100x with no visible change in results.

Reference graph

Works this paper leans on

42 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    J. D. Plummer and P. B. Griffin, Integrated Circuit Fabrication: Science and Technology. Cambridge University Press, 2023

    work page 2023

  2. [2]

    Green chemistry by nano-catalysis,

    V . Polshettiwar and R. S. Varma, “Green chemistry by nano-catalysis,” Green Chemistry, vol. 12, no. 5, pp. 743–754, 2010

    work page 2010

  3. [3]

    Nanomedicine—challenge and perspectives,

    K. Riehemann, S. W. Schneider, T. A. Luger, B. Godin, M. Ferrari, and H. Fuchs, “Nanomedicine—challenge and perspectives,” Angewandte Chemie International Edition , vol. 48, no. 5, pp. 872–897, 2009

    work page 2009

  4. [4]

    Energy storage: The future enabled by nanomaterials,

    E. Pomerantseva, F. Bonaccorso, X. Feng, Y . Cui, and Y . Gogotsi, “Energy storage: The future enabled by nanomaterials,” Science, vol. 366, no. 6468, p. eaan8285, 2019

    work page 2019

  5. [5]

    D. B. Williams, C. B. Carter, D. B. Williams, and C. B. Carter, The transmission electron microscope. Springer, 2009

    work page 2009

  6. [6]

    K2: A Super-Resolution Electron Counting Direct Detection Camera for Cryo-EM,

    C. Booth, “K2: A Super-Resolution Electron Counting Direct Detection Camera for Cryo-EM,” Microscopy and Microanalysis, vol. 18, no. S2, pp. 78–79, jul 1 2012

    work page 2012

  7. [7]

    High Dynamic Range Pixel Array Detector for Scanning Transmission Electron Microscopy,

    M. W. Tate, P. Purohit, D. Chamberlain, K. X. Nguyen, R. Hovden, C. S. Chang, P. Deb, E. Turgut, J. T. Heron, D. G. Schlom, D. C. Ralph, G. D. Fuchs, K. S. Shanks, H. T. Philipp, D. A. Muller, and S. M. Gruner, “High Dynamic Range Pixel Array Detector for Scanning Transmission Electron Microscopy,” Microscopy and Microanalysis , vol. 22, no. 1, pp. 237–24...

    work page 2016

  8. [8]

    Four-Dimensional Scanning Transmission Electron Mi- croscopy (4d-STEM): From Scanning Nanodiffraction to Ptychography and Beyond,

    C. Ophus, “Four-Dimensional Scanning Transmission Electron Mi- croscopy (4d-STEM): From Scanning Nanodiffraction to Ptychography and Beyond,” Microscopy and Microanalysis , vol. 25, no. 3, pp. 563– 582, 5 2019, publisher: Oxford University Press (OUP)

    work page 2019

Show all 42 references
  1. [9]

    Egami and S

    T. Egami and S. J. L. Billinge, Underneath the Bragg Peaks: Structural Analysis of Complex Materials . Newnes, dec 31 2012, google-Books- ID: kzeSEdYS54wC. 8

    2012

  2. [10]

    Mapping structure and morphology of amorphous organic thin films by 4D-STEM pair distribution function analysis,

    X. Mu, A. Mazilkin, C. Sprau, A. Colsmann, and C. K ¨ubel, “Mapping structure and morphology of amorphous organic thin films by 4D-STEM pair distribution function analysis,” Microscopy, vol. 68, no. 4, pp. 301– 309, 3 2019, publisher: Oxford University Press (OUP)

    2019

  3. [11]

    Cryogenic 4D- STEM analysis of an amorphous-crystalline polymer blend: Combined nanocrystalline and amorphous phase mapping,

    J. Donohue, S. E. Zeltmann, K. C. Bustillo, B. Savitzky, M. A. Jones, G. F. Meyers, C. Ophus, and A. M. Minor, “Cryogenic 4D- STEM analysis of an amorphous-crystalline polymer blend: Combined nanocrystalline and amorphous phase mapping,” iScience, vol. 25, no. 3, p. 103882, 3 ...

    2022

  4. [12]

    Learning the parts of objects by non- negative matrix factorization,

    D. D. Lee and H. S. Seung, “Learning the parts of objects by non- negative matrix factorization,” Nature, vol. 401, no. 6755, pp. 788–791, 10 1999, publisher: Springer Science and Business Media LLC

    1999

  5. [13]

    Fast Grain Mapping with Sub- Nanometer Resolution Using 4D-STEM with Grain Classification by Principal Component Analysis and Non-Negative Matrix Factorization,

    F. I. Allen, T. C. Pekin, A. Persaud, S. J. Rozeveld, G. F. Meyers, J. Ciston, C. Ophus, and A. M. Minor, “Fast Grain Mapping with Sub- Nanometer Resolution Using 4D-STEM with Grain Classification by Principal Component Analysis and Non-Negative Matrix Factorization,” Microsco...

    2021

  6. [14]

    Mapping local atomic structure of metallic glasses using ma- chine learning aided 4D-STEM,

    S. Kang, V . Wollersen, C. Minnert, K. Durst, H.-S. Kim, C. K ¨ubel, and X. Mu, “Mapping local atomic structure of metallic glasses using ma- chine learning aided 4D-STEM,” Acta Materialia, vol. 263, p. 119495, jan 15 2024

    2024

  7. [15]

    Non-negative matrix factorization for mining big data obtained using four-dimensional scanning transmission electron microscopy,

    F. Uesugi, S. Koshiya, J. Kikkawa, T. Nagai, K. Mitsuishi, and K. Ki- moto, “Non-negative matrix factorization for mining big data obtained using four-dimensional scanning transmission electron microscopy,” Ultramicroscopy, vol. 221, p. 113168, feb 1 2021

    2021

  8. [16]

    Unsupervised machine learning combined with 4D scanning transmission electron microscopy for bimodal nanostructural analysis,

    K. Kimoto, J. Kikkawa, K. Harano, O. Cretu, Y . Shibazaki, and F. Uesugi, “Unsupervised machine learning combined with 4D scanning transmission electron microscopy for bimodal nanostructural analysis,” Scientific Reports, vol. 14, no. 1, p. 2901, feb 5 2024, publisher: Nature ...

    2024

  9. [17]

    Randomized nonnegative matrix factorization,

    N. B. Erichson, A. Mendible, S. Wihlborn, and J. N. Kutz, “Randomized nonnegative matrix factorization,” Pattern Recognition Letters, vol. 104, pp. 1–7, 2018

    2018

  10. [18]

    Tio2 photocatalysis: Design and applica- tions,

    K. Nakata and A. Fujishima, “Tio2 photocatalysis: Design and applica- tions,” Journal of photochemistry and photobiology C: Photochemistry Reviews, vol. 13, no. 3, pp. 169–189, 2012

    2012

  11. [19]

    Interface-engineered amorphous tio2-based resistive memory devices,

    H. Y . Jeong, J. Y . Lee, and S.-Y . Choi, “Interface-engineered amorphous tio2-based resistive memory devices,” Advanced Functional Materials , vol. 20, no. 22, pp. 3912–3917, 2010

    2010

  12. [20]

    Understanding cathode–electrolyte interphase formation in solid state li-ion batteries via 4D-STEM,

    N. C. Paranamana, A. Werbrouck, A. K. Datta, X. He, and M. J. Young, “Understanding cathode–electrolyte interphase formation in solid state li-ion batteries via 4D-STEM,” Advanced Energy Materials , vol. 15, no. 11, p. 2403904, 2025

    2025

  13. [21]

    Cryo- ePDF: Overcoming Electron Beam Damage to Study the Local Atomic Structure of Amorphous ALD Aluminum Oxide Thin Films within a TEM,

    A. M. Jasim, X. He, Y . Xing, T. A. White, and M. J. Young, “Cryo- ePDF: Overcoming Electron Beam Damage to Study the Local Atomic Structure of Amorphous ALD Aluminum Oxide Thin Films within a TEM,” ACS Omega, vol. 6, no. 13, pp. 8986–9000, apr 6 2021

    2021

  14. [22]

    Growth of titanium dioxide thin films by atomic layer epitaxy,

    M. Ritala, M. Leskel ¨a, E. Nyk ¨anen, P. Soininen, and L. Niinist ¨o, “Growth of titanium dioxide thin films by atomic layer epitaxy,” Thin Solid Films, vol. 225, no. 1, pp. 288–295, mar 25 1993

    1993

  15. [23]

    Finding structure with randomness: Probabilistic algorithms for constructing approximate ma- trix decompositions,

    N. Halko, P.-G. Martinsson, and J. A. Tropp, “Finding structure with randomness: Probabilistic algorithms for constructing approximate ma- trix decompositions,” SIAM review, vol. 53, no. 2, pp. 217–288, 2011

    2011

  16. [24]

    Svd based initialization: A head start for nonnegative matrix factorization,

    C. Boutsidis and E. Gallopoulos, “Svd based initialization: A head start for nonnegative matrix factorization,”Pattern Recognition, vol. 41, no. 4, pp. 1350–1362, apr 1 2008

    2008

  17. [25]

    Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization,

    J. M. Maisog, A. T. DeMarco, K. Devarajan, S. S. Young, P. Fogel, and G. Luta, “Assessing Methods for Evaluating the Number of Components in Non-Negative Matrix Factorization,” Mathematics (Basel, Switzer- land), vol. 9, no. 22, p. 2840, nov 2 2021, pMID: 35694180 PMCID: PMC9181460

    2021

  18. [26]

    Cross-Validatory Estimation of the Number of Components in Factor and Principal Components Models,

    S. Wold, “Cross-Validatory Estimation of the Number of Components in Factor and Principal Components Models,” Technometrics, vol. 20, no. 4, pp. 397–405, nov 1 1978, publisher: ASA Website eprint: https://www.tandfonline.com/doi/pdf/10.1080/00401706.1978.10489693

    1978

  19. [27]

    Rank selection in low-rank matrix approximations: A study of cross-validation for nmfs,

    B. Kanagal and V . Sindhwani, “Rank selection in low-rank matrix approximations: A study of cross-validation for nmfs,” in Proc Conf Adv Neural Inf Process , vol. 1, 2010, pp. 10–15

    2010

  20. [28]

    Towards quantitative treatment of electron pair distribution function,

    T. E. Gorelik, R. Neder, M. W. Terban, Z. Lee, X. Mu, C. Jung, T. Jacob, and U. Kaiser, “Towards quantitative treatment of electron pair distribution function,” Acta Crystallographica Section B Structural Science, Crystal Engineering and Materials, vol. 75, no. 4, pp. 532–549,...

    2019

  21. [29]

    Probing the Atomic-Scale Structure of Amorphous Aluminum Oxide Grown by Atomic Layer Deposition,

    M. J. Young, N. M. Bedford, A. Yanguas-Gil, S. Letourneau, M. Coile, D. J. Mandia, B. Aoun, A. S. Cavanagh, S. M. George, and J. W. Elam, “Probing the Atomic-Scale Structure of Amorphous Aluminum Oxide Grown by Atomic Layer Deposition,” ACS Applied Materials & Interfaces, vol....

    2020

  22. [30]

    Measuring Local Atomic Structure Variations through the Depth of Ultrathin ( <20 nm) ALD Aluminum Oxide: Implications for Lithium-Ion Batteries,

    N. C. Paranamana, R. Gettler, H. Koenig, S. Montgomery-Smith, X. He, and M. J. Young, “Measuring Local Atomic Structure Variations through the Depth of Ultrathin ( <20 nm) ALD Aluminum Oxide: Implications for Lithium-Ion Batteries,” ACS Applied Nano Materials , vol. 5, no. 9, ...

    2022

  23. [31]

    Iterative reverse Monte Carlo and molecular statics for improved atomic structure modeling: a case study of zinc oxide grown by atomic layer deposition,

    R. C. Gettler, H. D. Koenig, and M. J. Young, “Iterative reverse Monte Carlo and molecular statics for improved atomic structure modeling: a case study of zinc oxide grown by atomic layer deposition,” Physical Chemistry Chemical Physics , vol. 23, no. 46, pp. 26 417–26 427, 2021

    2021

  24. [32]

    Simple ePDF: A Pair Distribution Function Method Based on Electron Diffraction Patterns to Reveal the Local Structure of Amorphous and Nanocrystalline Materials,

    J. L. L ´ab´ar, K. Hajagos-Nagy, P. P. Das, A. Gomez-Perez, and G. Radn ´oczi, “Simple ePDF: A Pair Distribution Function Method Based on Electron Diffraction Patterns to Reveal the Local Structure of Amorphous and Nanocrystalline Materials,” Nanomaterials, vol. 13, no. 24, p....

    2023

  25. [33]

    epdfpy: A Python-based interactive GUI tool for electron pair distri- bution function analysis of amorphous materials,

    M. Kim, P. Kim, R. Bassiri, K. Prasai, M. M. Fejer, and K.-h. Lee, “epdfpy: A Python-based interactive GUI tool for electron pair distri- bution function analysis of amorphous materials,” Computer Physics Communications, vol. 299, p. 109137, jun 1 2024

    2024

  26. [34]

    Pdfgetx3: a rapid and highly automatable program for processing powder diffraction data into total scattering pair distribution functions,

    P. Juh ´as, T. Davis, C. L. Farrow, and S. J. L. Billinge, “Pdfgetx3: a rapid and highly automatable program for processing powder diffraction data into total scattering pair distribution functions,” Journal of Applied Crys- tallography, vol. 46, no. 2, pp. 560–566, 4 2013, pu...

    2013

  27. [35]

    The fast azimuthal integration Python library: pyfai,

    G. Ashiotis, A. Deschildre, Z. Nawaz, J. P. Wright, D. Karkoulis, F. E. Picca, and J. Kieffer, “The fast azimuthal integration Python library: pyfai,” Journal of Applied Crystallography, vol. 48, no. 2, pp. 510–519, apr 1 2015, publisher: International Union of Crystallography

    2015

  28. [36]

    hyperspy/hyperspy: v2.1.0,

    Francisco de la Pe ˜na, Eric Prestat, Vidar Tonaas Fauske, Jonas L¨ahnemann, Pierre Burdet, Petras Jokubauskas, Tom Furnival, Carter Francis, Magnus Nord, Tomas Ostasevicius, Katherine E. MacArthur, Duncan N. Johnstone, Mike Sarahan, Joshua Taillon, Thomas Aarholt, pquinn-dls,...

    2024

  29. [37]

    openNCEM documentation,

    “openNCEM documentation,” https://openncem.readthedocs.io/en/latest/#, [Online; accessed 2025-07-03]. [Online]. Available: https://openncem.readthedocs.io/en/latest/#

    2025

  30. [38]

    nmfmapping: a cloud-based web application for non-negative matrix factorization of powder diffraction and pair distribution function datasets,

    Z. Thatcher, C.-H. Liu, L. Yang, B. C. McBride, G. T. Tran, A. Wus- trow, M. A. Karlsen, J. R. Neilson, D. B. Ravnsbæk, and S. J. L. Billinge, “nmfmapping: a cloud-based web application for non-negative matrix factorization of powder diffraction and pair distribution function ...

    2022

  31. [39]

    A fast two-stage algorithm for non- negative matrix factorization in smoothly varying data,

    R. Gu, S. J. Billinge, and Q. Du, “A fast two-stage algorithm for non- negative matrix factorization in smoothly varying data,” Foundations of Crystallography, vol. 79, no. 2, pp. 203–216, 2023

    2023

  32. [40]

    Validation of non-negative matrix factorization for rapid assessment of large sets of atomic pair distribution function data,

    C.-H. Liu, C. J. Wright, R. Gu, S. Bandi, A. Wustrow, P. K. Todd, D. O’Nolan, M. L. Beauvais, J. R. Neilson, P. J. Chupas, K. W. Chapman, and S. J. L. Billinge, “Validation of non-negative matrix factorization for rapid assessment of large sets of atomic pair distribution func...

    2021

  33. [41]

    Mapping structural heterogeneity at the nanoscale with scanning nano- structure electron microscopy (SNEM),

    Y . Rakita, J. L. Hart, P. P. Das, S. Shahrezaei, D. L. Foley, S. N. Mathaudhu, S. Nicolopoulos, M. L. Taheri, and S. J. L. Billinge, “Mapping structural heterogeneity at the nanoscale with scanning nano- structure electron microscopy (SNEM),” Acta Materialia , vol. 242, p. 11...

    2023

  34. [42]

    Unveiling local atomic bonding and packing of amorphous nanophases via independent component analysis facilitated pair distribution function,

    X. Mu, L. Chen, R. Mikut, H. Hahn, and C. K ¨ubel, “Unveiling local atomic bonding and packing of amorphous nanophases via independent component analysis facilitated pair distribution function,” Acta Materi- alia, vol. 212, p. 116932, jun 15 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.