Pith. sign in

REVIEW 3 major objections 6 minor 24 references

Defect complexes in CrSBr revealed through electron microscopy and deep learning

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Bilayer CrSBr hosts a catalog of atomic defect complexes—Cr vacancy plus Cr interstitial, mixed Cr/Br vacancies, and 1D vacancy chains—whose occurrence matches computed binding-energy densities.

desk verdict Real new defect physics in CrSBr, honestly reported, but the headline numbers lean on a detector with no experimental ground truth — worth refereeing, needs a quantitative humility pass. read the letter →

arxiv 2506.08100 v1 pith:65OEXXCO submitted 2025-06-09 cond-mat.mtrl-sci cond-mat.mes-hall

Mads Weile , Sergii Grytsiuk , Aubrey Penn , Daniel G. Chica , Xavier Roy , Kseniia Mosina , Zdenek Sofer , Jakob Schi{o}tz
show 4 more authors
Stig Helveg Malte Rösner Frances M. Ross Julian Klein
This is my paper
classification cond-mat.mtrl-scicond-mat.mes-hall
keywords defectcomplexesCrSBrscanningtransmissionelectronmicroscopydeeplearningvanderWaalsmaterialsDFT+Uvacancychainssingle-photonemitters
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 sets out to build a comprehensive atomic-defect library for bilayer CrSBr, a magnetic quasi-1D van der Waals semiconductor, by combining scanning transmission electron microscopy with a deep-learning detector trained on simulated images. The claim is that what look like single point vacancies are often composite defects: Cr vacancies paired with Cr interstitials at specific neighboring sites, Cr vacancies paired with one or two Br vacancies, and extended 1D chains of Br vacancies along the a-axis with a power-law length distribution. The authors report that the relative occurrence of these complexes matches their DFT+U binding energy densities, with the most abundant interstitial complex VCr+Crδ2int having the lowest binding energy density (105 meV), and that DFT+U predicts highly localized, optically active in-gap states for the interstitial complexes. A sympathetic reader should care because this would be the first atomic-scale defect census for the FeOCl structural family, and it suggests concrete candidate defect emitters and a microscopic origin for CrSBr's unexplained transport anisotropy.

What carries the argument

The load-bearing machinery is threefold. First, an ensemble of four deep convolutional neural networks (UNet++, a nested U-Net segmentation architecture) is trained on hundreds of simulated HAADF-STEM images of 2L CrSBr flakes, augmented with realistic noise including fractal surface contamination, to detect and classify vacancy defects, while a second 'atom spotter' network locates atom columns with few-pm precision to align class averages. Second, class-averaging of the detected defects is the statistical tool that separates faint, real neighbor-column intensity changes from noise and thereby exposes the interstitial and vacancy complexes. Third, spin-resolved DFT+U calculations in 4×3×2 supercells provide the energetic ordering, expressed as binding energy densities (for example Ebind = (EN − (EN−d + Ed))/Ncells for removal of a Cr interstitial), which is the quantity the experimental occurrence is compared against. The quantitative anchor of the whole argument is the pairing of observed occurrence with Ebind: the most frequently observed interstitial complex has the lowest binding energy density, and the Cr/Br vacancy complexes are stabilized relative to an isolated Cr vacancy.

What would settle it

Take images of a fresh CrSBr bilayer with a beam energy below the Cr displacement threshold (for instance 60 keV) and count the relative abundance of VCr+Crδ2int and VCr+Crδ1int; if the ratio changes or the complexes vanish relative to 200 keV imaging, the complexes are beam products, not equilibrium motifs, and the claimed match to binding energy densities would not certify intrinsic defect physics.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that 2L CrSBr contains a library of defect complexes that are resolvable and quantifiable by a machine-learning-assisted HAADF-STEM workflow, and that their abundances are thermodynamically ordered by ab initio binding energy densities. Class-averaging of detected Cr vacancies reveals two families: Cr interstitials that vacate the Cr site and settle in specific nearby columns (VCr+Crδnint, dominated by the δ2 and δ1 positions), and combined Cr/Br vacancies (VCr+VBr, VCr+V2Br) that follow from the exclusive Cr–Br bonding of the structure. The same analysis exposes 1D chains of Br vacancies (up to 7 vacancies long, more than 2 nm) whose length distribution is best fit by a power law, interpreted as correlated, self-propagating growth. The authors show that simulated images of DFT+U relaxed structures reproduce the experimental class averages, that the observed ordering of complex occurrence agrees with computed binding energy densities, and that these complexes create multiple weakly dispersive in-gap defect states, with the interstitial complexes predicted to be optically active and spin-polarized according to the hosting monolayer.

Load-bearing premise

The central claim rests on the detector, trained only on simulated images, finding real defects in experimental images without bias toward one defect type; the paper's own supplement says that getting a true measure of the detector's performance is 'a clear challenge.'

Editorial extensions

If this is right

  • The most abundant interstitial complex, VCr+Crδ2int, is a real, structurally resolved configuration of bilayer CrSBr whose frequency is ordered by its low binding energy density (105 meV).
  • The interstitial and mixed Cr/Br vacancy complexes introduce multiple weakly dispersive in-gap states with both occupied and unoccupied levels, making them concrete candidate single-photon emitters with spin character inherited from the host monolayer.
  • The 1D Br-vacancy chains follow a (truncated) power-law length distribution with exponent α≈3.29, implying correlated growth rather than independent vacancy formation; DFT-computed infinite line defects are metallic, giving a microscopic mechanism for the gate-dependent conductivity anisotropy of CrSBr.
  • The same defect complexes are expected in more than 20 other FeOCl-type materials (transition-metal oxyhalides and chalcohalides), so the library is a blueprint for defect studies there.
  • The detector workflow—simulated training, ensemble detection, class averaging—extends to other beam-sensitive multilayer van der Waals materials and could transfer to imaging modalities such as ptychography.

Reading between the lines

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

  • The reported absolute defect densities (e.g., 1.5×10^14 cm⁻² for VBr) should be read as estimates rather than intrinsic concentrations, since the paper says they substantially exceed intrinsic values and partly reflect sample transfer and false positives; the intrinsic defect physics of pristine CrSBr remains untested.
  • If the large VBr line-length exponent (α≈3.29) is confirmed on larger fields of view, it would indicate a correlated-growth process with a finite cutoff; a null model of randomly placed independent vacancies would instead produce an exponential length distribution, so the power law is directly testable.
  • The predicted optical activity is probeable by correlating defect density with micro-photoluminescence intensity or by magneto-PL: if the in-gap states inherit the monolayer spin, defect emission lines should shift or split across the magnetic transition.
  • A defect-type-specific error analysis on simulated images with experimental noise statistics would quantify how much the 43% and 19% co-occurrence fractions and the line-length power law are biased by detection thresholds; the authors' thresholds (0.95 for VBr, 0.90 otherwise) make VBr counts the most uncertain.
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

3 major / 6 minor

Summary. The manuscript reports an experimental defect library for bilayer CrSBr obtained by combining HAADF-STEM imaging, a UNet++-based deep-learning detection workflow trained on simulated images, class averaging, and DFT+U calculations. The authors identify single Cr and Br vacancies, vertically stacked bi-vacancies, Cr-interstitial–vacancy complexes (VCr+Crδ1_int and VCr+Crδ2_int), Cr–Br vacancy complexes (VCr+VBr and VCr+V2Br), and extended 1D Br-vacancy line defects along the a-axis. They report defect densities, 43% and 19% co-occurrence fractions for VCr with VBr and V2Br, a truncated power-law distribution of Br-vacancy line lengths, and DFT binding-energy densities that rank VCr+Crδ2_int as the most stable interstitial complex. The paper also presents spin-resolved DFT+U band structures and defect charge densities, arguing that several complexes create highly localized, possibly optically active in-gap states. The central quantitative claim is that observed defect-complex occurrence agrees with the DFT binding-energy ordering, with the paper stating for VCr+Crδ2_int: 'Its relatively frequent occurrence is in excellent agreement with the lowest binding energy density Ebindδ2=105 meV.'

Significance. If the central claim holds, the paper provides a valuable defect library for a technologically important van der Waals magnet and outlines a transferable simulated-data-to-experiment deep-learning workflow for beam-sensitive materials. The DFT calculations are independent of the microscopy analysis, the image-simulation parameters and code are provided, and the class averages with simulated-image comparison give reasonable qualitative support for the identity of the single vacancies and the defect complexes. The paper also makes a falsifiable prediction about localized, optically active in-gap states. However, the quantitative occurrence statistics—the relative abundances, the co-occurrence fractions, and the line-length power law—rest on the assumption that the DCNN transfers from purely synthetic training data to experimental images with defect-type-independent precision and recall. That assumption is not validated with experimental ground truth, and the manuscript itself notes the difficulty of getting 'a true sense of their performance.' The significance of the qualitative defect library is real, but the quantitative agreement with DFT is not yet established.

major comments (3)
  1. [SM Sec. 4.7 and main text Methods] The DCNN is trained exclusively on simulated images with synthetic Perlin noise, and the only experimental validation (SM Fig. S12a-b) measures ensemble agreement and intensity histograms, not per-defect precision/recall. Because the detection threshold is 0.95 for VBr and 0.90 for all other defects, and because VBr has the weakest contrast change relative to its pristine column, the false-positive and false-negative rates can differ substantially by defect type. This directly biases the claimed relative abundances (VBr most abundant; VCr+Crδ2_int most common complex), the 43%/19% VCr–VBr co-occurrence fractions, and the Br-vacancy line-length distribution. Please report per-class precision and recall on a manually labeled subset of experimental images, or otherwise quantify the defect-type-dependent transfer gap; without this, the quantitative occurrence agreement with DFT binding-energy densities is not established.
  2. [Fig. 3 and Table I] The class averages for VCr+Crδ2_int, VCr+Crδ1_int, VCr+VBr, and VCr+V2Br contain N=11, 3, 12, and 6 images, respectively, and the complexes are identified by manual search through all detected VCr (main text: 'Manually searching through images of all detected VCr...'). The claim that VCr+Crδ2_int is 'most common' and 'in excellent agreement with' Ebind=105 meV is therefore based on a small, manually selected sample with no stated selection criteria or inter-rater reliability. Please provide the selection criteria, the total number of candidate images examined, and an automated or blinded counting procedure, or downgrade the quantitative ranking claim to a qualitative identification.
  3. [SM Sec. 7 and Fig. 4] The power-law fit for Br-vacancy line lengths is presented as evidence of correlated growth, but the authors themselves note that false negatives split long lines into shorter ones and false positives increase short-line counts. The truncated power law with α=3.29 is fitted without propagating these detection errors. A Monte Carlo simulation of the detection process, using estimated per-defect recall and false-positive rates, should be used to show that the observed length distribution is robust; otherwise the correlated-growth conclusion is not quantitatively supported.
minor comments (6)
  1. [Throughout] The notation for vertically stacked vacancies is inconsistent: Fig. 2 uses '2V_bi_Br' and '2V_bi_Cr' while the text also writes 'V_bi_2Br' and 'V_bi_2Cr'. Please unify the notation.
  2. [SM Fig. S2 caption] The beam current is given as '35 A' in the caption; it should be '35 pA'.
  3. [Main text, line-defect paragraph] The text contains 'hydrolysis from H20 or oxidation from O2'; 'H20' should be 'H2O'.
  4. [Fig. 4 and SM Sec. 7] Please state the fit parameters (α, truncation length, and uncertainty) and the fit range directly in the figure or caption, and specify which defect-line configurations are included in the experimental counts.
  5. [Fig. 3(j) caption] The caption says line profiles are compared along the b- or a-direction; please indicate which panel corresponds to which direction so readers can assess the comparison.
  6. [References] The GitHub repository is cited as a URL; please provide a versioned DOI or release identifier to ensure reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

No parameter was fitted to enforce agreement; the only mild circularity is a simulation-based training/validation consistency loop, not a derivation-by-construction.

  1. other [Methods (ML workflow) and Results (Fig. 2c-d, Fig. 3i-j); SM Sec. 4]
    "For the machine learning workflow and training of our DCNNs, we simulate 600 HAADF-STEM images of 30x30 Å2 flakes of CrSBr... We compare the experimental images with simulations (see Fig. 2(c)) that are based on DFT relaxed crystal structures."

    The defect detector is trained exclusively on simulated HAADF-STEM images of DFT-relaxed structures (600 images, 40 augmentations each; SM Sec. 4), and the same simulation pipeline (abTEM/PRISM with DFT-relaxed structures) is then used as the reference for the 'excellent agreement' of the experimental class averages (Fig. 2c-d and Fig. 3i-j). This creates a consistency loop: the network has learned the contrast signatures of the very simulation models used to validate the identifications. The loop is not a strict derivation-by-construction because the class averages are measured intensities, the DFT binding energies are independent of the measured frequencies, and the defect complexes were not training classes; hence this is a minor validation bias rather than a forced reduction.

full rationale

The paper's central claim—that observed relative occurrence of VCr+Crδ2_int over VCr+Crδ1_int agrees with DFT+U binding energy densities—does not reduce to its inputs by construction. The binding energies E_bind are computed ab initio from relaxed supercells and are never fitted to the experimental defect counts; the defect complexes are not among the DCNN training classes; and the 43%/19% co-occurrence fractions and line-length distribution are measured statistics, not outputs of the model. The only mild circularity is methodological: the detector is trained exclusively on simulated HAADF-STEM images of DFT-relaxed structures, and the same simulation pipeline is then used as the reference for 'excellent agreement' of experimental class averages. This is a consistency loop that can bias validation, but it does not force the central result because the class averages are measured intensities and the DFT energetics are independent of the measured frequencies. Self-citations (e.g., U and J from prior cRPA work by co-author Rösner; prior CrSBr electron-beam studies) are contextual or constitute independent cRPA parameter calculations with stated assumptions, not load-bearing circular arguments. The DCNN transferability concern is a correctness/robustness risk, not a circularity. Score 2 reflects the minor simulation-based validation loop.

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

No new particles, forces, or dimensions are introduced. The evidential load is carried by DFT+U (with U and J as chosen parameters), PRISM image simulation, and the assumed transferability of a synthetic-data-trained DCNN; these are listed explicitly above.

free parameters (4)
  • Hubbard U for Cr d orbitals = 2.5 eV
    On-site Coulomb interaction from constrained RPA for monolayer CrSBr in a dielectric environment; it enters every DFT+U structure relaxation, binding energy, and band structure, and is not fitted to the defect data in this paper.
  • Hund coupling J for Cr d orbitals = 0.4 eV
    Exchange parameter paired with U, also from prior cRPA work; affects magnetic and electronic structure of all computed defect configurations.
  • DCNN detection thresholds = 0.95 for VBr; 0.90 for other defects
    Chosen by hand to reduce false positives; directly controls all defect counts, class-average compositions, and the Br vacancy line-length distribution.
  • Truncated power-law exponent = alpha = 3.29
    Obtained by fitting a truncated power law P(x)=x^-alpha e^{-x/x_max} to five binned Br vacancy line-length counts; used to support correlated line growth.
assumptions (4)
  • domain assumption GGA+U with U=2.5 eV, J=0.4 eV accurately describes relaxed geometries, binding energies, and in-gap states of defective CrSBr.
    All comparisons of observed versus computed defect occurrence and all electronic structure claims depend on this approximation; the pristine gap is underestimated, as the authors note.
  • domain assumption PRISM/abTEM HAADF-STEM simulations with the stated aberrations and 80 pm Gaussian smoothing reproduce experimental imaging contrast for 2L CrSBr.
    The same simulation approach generates both the DCNN training data and the simulated class averages used to validate experimental observations.
  • domain assumption A DCNN trained exclusively on simulated images generalizes to experimental HAADF-STEM images.
    The SM states that obtaining a true sense of performance is a key challenge and validation relies on ensemble consistency and column-intensity distributions rather than atom-by-atom experimental labels.
  • domain assumption Co-existence of occupied and unoccupied in-gap DFT states implies optically active defect transitions.
    Used to label defects as optically active and to suggest single-photon emitters; optical matrix elements, radiative rates, and excitation energies are not computed or measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Defect complexes in CrSBr revealed through electron microscopy and deep learning." pith.science (2026). https://pith.science/paper/65OEXXCO

@misc{pith2026250608100,
  author       = {Pith},
  title        = {Pith review of: Defect complexes in CrSBr revealed through electron microscopy and deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65OEXXCO}},
  note         = {Machine review of arXiv:2506.08100}
}
read the original abstract

Atomic defects underpin the properties of van der Waals materials, and their understanding is essential for advancing quantum and energy technologies. Scanning transmission electron microscopy is a powerful tool for defect identification in atomically thin materials, and extending it to multilayer and beam-sensitive materials would accelerate their exploration. Here we establish a comprehensive defect library in a bilayer of the magnetic quasi-1D semiconductor CrSBr by combining atomic-resolution imaging, deep learning, and ab-initio calculations. We apply a custom-developed machine learning work flow to detect, classify and average point vacancy defects. This classification enables us to uncover several distinct Cr interstitial defect complexes, combined Cr and Br vacancy defect complexes and lines of vacancy defects that extend over many unit cells. We show that their occurrence is in agreement with our computed structures and binding energy densities, reflecting the intriguing layer interlocked crystal structure of CrSBr. Our ab-initio calculations show that the interstitial defect complexes give rise to highly localized electronic states. These states are of particular interest due to the reduced electronic dimensionality and magnetic properties of CrSBr and are furthermore predicted to be optically active. Our results broaden the scope of defect studies in challenging materials and reveal new defect types in bilayer CrSBr that can be extrapolated to the bulk and to over 20 materials belonging to the same FeOCl structural family.

Figures

Figures reproduced from arXiv: 2506.08100 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 10 canonical work pages

  1. [1]

    G. F. Schneider, V. E. Calado, H. Zandbergen, L. M. K. Vandersypen, and C. Dekker, Nano Letters 10, 1912–1916 (2010), ISSN 1530-6992, URL http://dx.doi.org/10.1021/nl1008037

    work page doi:10.1021/nl1008037 2010

  2. [2]

    Torres, A

    K. Torres, A. Kuc, L. Maschio, T. Pham, K. Reidy, L. Dekanovsky, Z. Sofer, F. M. Ross, and J. Klein, Advanced Functional Materials 33 (2023), URL https://doi.org/10.1002/adfm.202211366

    doi:10.1002/adfm.202211366 2023

  3. [3]

    Klein, T

    J. Klein, T. Pham, J. D. Thomsen, J. B. Curtis, T. Denneulin, M. Lorke, M. Florian, A. Steinhoff, R. A. Wiscons, J. Luxa, et al., Nature Communications 13 (2022), URL https://doi.org/10.1038/s41467-022-32737-8

    doi:10.1038/s41467-022-32737-8 2022

  4. [4]

    A Chat with Andrew on MLOps: From Model-centric to Data-centric AI , URL https://www.youtube.com/watch?v= 06-AZXmwHjo

    work page

  5. [5]

    Ghosh, B

    A. Ghosh, B. G. Sumpter, O. Dyck, S. V. Kalinin, and M. Ziatdinov, npj Computational Materials 7 (2021), ISSN 2057- 3960, URL http://dx.doi.org/10.1038/s41524-021-00569-7

    doi:10.1038/s41524-021-00569-7 2021

  6. [6]

    Madsen, P

    J. Madsen, P. Liu, J. Kling, J. B. Wagner, T. W. Hansen, O. Winther, and J. Schiøtz, Advanced Theory and Simulations 1, 1800037 (2018), ISSN 2513-0390

    work page 2018

  7. [7]

    Deep Learning of Atomically Resolved Scanning Transmission Electron Microscopy Images: Chemical Identification and Tracking Local Transformations

    M. Ziatdinov, O. Dyck, A. Maksov, X. Li, X. Sang, K. Xiao, R. R. Unocic, R. Vasudevan, S. Jesse, and S. V. Kalinin, ACS Nano 11, 12742 (2017), ISSN 1936-0851, 1936-086X, 1801.05860

    work page Pith review arXiv 2017

  8. [8]

    Hjorth Larsen, J

    A. Hjorth Larsen, J. Jørgen Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, et al., Journal of Physics: Condensed Matter 29, 273002 (2017), ISSN 1361-648X, URL http: //dx.doi.org/10.1088/1361-648X/aa680e

    doi:10.1088/1361-648x/aa680e 2017

Show all 24 references
  1. [9]

    Madsen and T

    J. Madsen and T. Susi, Open Research Europe 1, 24 (2021), ISSN 2732-5121, URL http://dx.doi.org/10.12688/ openreseurope.13015.2

    2021

  2. [10]

    Ophus, Advanced Structural and Chemical Imaging 3 (2017), ISSN 2198-0926, URL http://dx.doi.org/10.1186/ s40679-017-0046-1

    C. Ophus, Advanced Structural and Chemical Imaging 3 (2017), ISSN 2198-0926, URL http://dx.doi.org/10.1186/ s40679-017-0046-1

    2017

  3. [11]

    Lin, C.-C

    Y.-C. Lin, C.-C. Lu, C.-H. Yeh, C. Jin, K. Suenaga, and P.-W. Chiu, Nano Letters 12, 414–419 (2011), ISSN 1530-6992, URL http://dx.doi.org/10.1021/nl203733r

    2011 doi

  4. [12]

    O. Dyck, A. R. Lupini, P. D. Rack, J. Fowlkes, and S. Jesse, Nano Select 3, 643–654 (2021), ISSN 2688-4011, URL http://dx.doi.org/10.1002/nano.202100188

    2021 doi

  5. [13]

    O. Dyck, S. Kim, S. V. Kalinin, and S. Jesse, Journal of Vacuum Science & Technology B, Nanotechnology and Microelec- tronics: Materials, Processing, Measurement, and Phenomena 36 (2017), ISSN 2166-2754, URL http://dx.doi.org/10. 1116/1.5003034

    2017

  6. [14]

    R. Lin, R. Zhang, C. Wang, X.-Q. Yang, and H. L. Xin, Scientific Reports 11 (2021), ISSN 2045-2322, URL http: //dx.doi.org/10.1038/s41598-021-84499-w

    2021 doi

  7. [15]

    C.-H. Lee, A. Khan, D. Luo, T. P. Santos, C. Shi, B. E. Janicek, S. Kang, W. Zhu, N. A. Sobh, A. Schleife, et al., Nano Letters 20, 3369–3377 (2020), ISSN 1530-6992, URL http://dx.doi.org/10.1021/acs.nanolett.0c00269

    2020 doi

  8. [16]

    K. Lee, J. Park, S. Choi, Y. Lee, S. Lee, J. Jung, J.-Y. Lee, F. Ullah, Z. Tahir, Y. S. Kim, et al., Nano Letters 22, 4677–4685 (2022), ISSN 1530-6992, URL http://dx.doi.org/10.1021/acs.nanolett.2c00550

    2022 doi

  9. [17]

    Eevee, Perlin noise (2016), accessed: 2022-05-31, URL https://eev.ee/blog/2016/05/29/perlin-noise/

    2016

  10. [18]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, in 2009 IEEE Conference on Computer Vision and Pattern 31 Recognition (IEEE, 2009), URL http://dx.doi.org/10.1109/CVPR.2009.5206848

    2009

  11. [19]

    Reuther, J

    A. Reuther, J. Kepner, C. Byun, S. Samsi, W. Arcand, D. Bestor, B. Bergeron, V. Gadepally, M. Houle, M. Hubbell, et al., in 2018 IEEE High Performance extreme Computing Conference (HPEC) (IEEE, 2018), URL http://dx.doi.org/ 10.1109/HPEC.2018.8547629

    2018

  12. [20]

    Ziatdinov, O

    M. Ziatdinov, O. Dyck, X. Li, B. G. Sumpter, S. Jesse, R. K. Vasudevan, and S. V. Kalinin, Science Advances 5 (2019), ISSN 2375-2548, URL http://dx.doi.org/10.1126/sciadv.aaw8989

    2019 doi

  13. [21]

    Klein, Z

    J. Klein, Z. Song, B. Pingault, F. Dirnberger, H. Chi, J. B. Curtis, R. Dana, R. Bushati, J. Quan, L. Dekanovsky, et al., ACS Nano 17, 288 (2022), URL https://doi.org/10.1021/acsnano.2c07655

    2022 doi

  14. [22]

    Iakubovskii, Segmentation models pytorch , https://github.com/qubvel/segmentation_models.pytorch (2019)

    P. Iakubovskii, Segmentation models pytorch , https://github.com/qubvel/segmentation_models.pytorch (2019)

    2019

  15. [23]

    F. Wu, I. Gutiérrez-Lezama, S. A. López-Paz, M. Gibertini, K. Watanabe, T. Taniguchi, F. O. von Rohr, N. Ubrig, and A. F. Morpurgo, Advanced Materials 34, 2109759 (2022), URL https://doi.org/10.1002/adma.202109759

    2022 doi

  16. [24]

    E. J. Telford, A. H. Dismukes, K. Lee, M. Cheng, A. Wieteska, A. K. Bartholomew, Y.-S. Chen, X. Xu, A. N. Pasupathy, X. Zhu, et al., Advanced Materials 32, 2003240 (2020), URL https://doi.org/10.1002/adma.202003240. 32

    2020 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.