Pith. sign in

REVIEW 3 major objections 4 minor 57 references

Machine learning of microstructure--property relationships in materials leveraging microstructure representation from foundational vision transformers

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Pre-trained vision transformers describe microstructures well enough to predict material properties without materials-specific training.

desk verdict Useful, honest application paper with a credible negative result in the simulation case; the experimental hardness win is real-looking but rests on ungrouped cross-validation, so the 13.5% vs 17.9% gap is provisional. read the letter →

arxiv 2501.18637 v2 pith:IUDIX6BN submitted 2025-01-28 cs.CV cond-mat.mtrl-scics.LGphysics.comp-ph

Sheila E. Whitman , Marat I. Latypov This is my paper · ORCID
classification cs.CVcond-mat.mtrl-scics.LGphysics.comp-ph
keywords microstructure–propertyrelationshipsmicrostructurerepresentationvisiontransformerfoundationalmodelsmachinelearningmaterialsinformaticstwo-pointcorrelationssuperalloyhardness
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 proposes that general-purpose visual features extracted by pre-trained vision transformers (CLIP, DINOv2, SAM) can serve as a task-agnostic microstructure representation, enabling machine learning of microstructure–property relationships without training or fine-tuning the transformers on any materials data. If true, materials scientists could build property predictors directly from raw microstructure images, bypassing the labor-intensive segmentation and descriptor selection used in traditional approaches. The authors test the hypothesis in two case studies: predicting Young's modulus of synthetic two-phase microstructures from three orthogonal 2D sections, and predicting Vickers hardness of Ni- and Co-base superalloys from published SEM images. The experimental case shows the best ViT-based model (DINOv2 features with support vector regression) reaching a mean absolute percentage error of 13.5%, compared with 17.9% for two-point correlations and 17.2% for a domain-specific CNN, suggesting the approach is most advantageous when raw images are noisy and segmentation is error-prone.

What carries the argument

The central object is a pre-trained vision transformer used as a fixed feature extractor. For CLIP and DINOv2, the image-level [CLS] token, a vector that summarizes the whole image, is taken directly as the microstructure descriptor; for SAM, which lacks such a token, patch-level features are aggregated. The extracted high-dimensional vectors are reduced by principal component analysis and then fed to simple regression models. This machinery replaces the usual pipeline of hand-picking geometric or statistical microstructure descriptors (or training a bespoke CNN) with a single forward pass through a model whose weights were learned on natural images, and the comparison to two-point correlations and a domain-specific CNN is what carries the evidence for the claim.

What would settle it

Take a dataset of micrographs and hardness values collected under a single standardized testing protocol on identically prepared samples, then compare the best ViT-based model (DINOv2 features with SVR) against two-point correlations computed on carefully segmented versions. If the ViT advantage shrinks or disappears under these controlled conditions, the reported 13.5% versus 17.9% margin would be attributable to dataset heterogeneity rather than to the feature representation itself.

Watch

Extended reading notes

Core claim

The central claim is the hypothesis, stated in the introduction, that the general-purpose visual features pre-trained ViTs extract from images can serve as robust microstructure representation for modeling properties without training or fine-tuning the ViTs to any materials data. The authors demonstrate this by freezing the weights of CLIP, DINOv2, and SAM, taking their image-level (or aggregated patch-level) feature vectors, reducing them with PCA, and training lightweight linear, polynomial, or support-vector regressors. In the simulation-based case, features from these ViTs yielded accuracy comparable to the traditional two-point correlation descriptor (best MAPE 25.1% vs 24.1%), while in the experimental hardness case all ViT-based models outperformed corresponding two-point-correlation models, with the best DINOv2-SVR model reaching 13.5% mean MAPE. The paper also finds that concatenating features from multiple 2D sections is consistently better than mean pooling, and that the principal components of ViT features are less dominated by phase volume fraction than those of two-point correlations, suggesting they preserve finer microstructural detail.

Load-bearing premise

The pooled experimental hardness values from 19 papers are treated as directly comparable even though they come from different testing loads, sample preparations, and image regions that may not match the indented area; if those measurements are not commensurable, the reported accuracy of every compared method is affected by uncontrolled noise.

Editorial extensions

If this is right

  • If the hypothesis holds, property prediction from raw microstructure images no longer requires phase segmentation, a step that is sensitive to imaging conditions and prone to errors on experimental micrographs.
  • Off-the-shelf ViT features could serve as a common microstructure representation across many material classes and properties, replacing task-specific descriptor selection.
  • The approach lowers the computational barrier: feature extraction on 5900 microstructures took minutes to hours on a consumer laptop, compared with 48 hours of GPU training reported for a task-specific CNN.
  • Adding alloy composition to the ViT features yields only marginal improvement (0.9% MAPE) for superalloy hardness, hinting that microstructure features may already encode much of the compositional effect, though the paper leaves universality open.

Reading between the lines

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

  • The segmentation-free advantage suggests a testable extension: apply the same framework to other properties where segmentation is known to be unreliable and check whether the ViT margin over two-point correlations grows with segmentation difficulty.
  • If ViT features encode composition implicitly, a sharper test would be to hold microstructure constant while varying composition in synthetic images; the model should be insensitive to compositional changes if it relies only on visible structure, which is not automatically guaranteed.
  • The finding that ViT principal components are less volume-fraction-dominated points toward using these features for properties sensitive to morphology and spatial arrangement, where volume-fraction-heavy descriptors underperform; this is our inference, not tested in the paper.
  • The results on small experimental datasets suggest foundation-model features may be especially valuable in data-scarce materials domains; a systematic study of accuracy versus training-set size would quantify this benefit.
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 / 4 minor

Summary. The paper proposes using pre-trained vision transformers (CLIP, DINOv2, SAM) as frozen feature extractors for microstructure images, followed by PCA dimensionality reduction and lightweight regression, to predict material properties without training or fine-tuning the transformer on materials data. The approach is evaluated in two case studies: Young's modulus of simulated two-phase microstructures from three orthogonal 2D sections, and Vickers hardness of Ni- and Co-base superalloys from 149 experimental SEM images pooled from 19 papers. In the first case study, ViT-based polynomial regression achieves MAPE comparable to two-point correlations (25.1% versus 24.1% for the best models); in the second, DINOv2 features with SVR achieve 13.5% MAPE versus 17.9% for two-point correlations and 17.2% for a domain CNN. The paper concludes that foundational ViT features provide a robust, segmentation-free microstructure representation for materials data with scarce labels, and it also explores appending composition vectors, which gives a modest further improvement.

Significance. If the case-study results hold, the approach is genuinely useful: it offers a training-free, segmentation-free descriptor that works on raw micrographs, is computationally light at inference time, and is backed by released code. The study is honest in reporting that ViT features do not beat two-point correlations in the simulated binary-image case, and the comparison against a domain-specific CNN and two-point correlations is a strength. However, the central claim rests heavily on the second case study, whose evaluation has a source-grouping leakage risk: the ungrouped 10-fold cross-validation allows images from the same source paper in both training and test folds, so the reported 13.5% versus 17.9% advantage may be inflated by paper-specific image and label artifacts. The significance is therefore conditional on a grouped or leave-one-paper-out analysis being performed and on the commensurability of the hardness labels being addressed.

major comments (3)
  1. [Section 3.2, Figure 8] The nested 10-fold cross-validation is performed on ungrouped samples, so micrographs from the same source paper can appear in both training and test folds. Because the hardness values are pooled from 19 papers without harmonizing indentation loads or verifying that the imaged area coincides with the indented region, a model can exploit paper-specific image statistics as a shortcut. The reported 13.5% versus 17.9% MAPE advantage of DINOv2 over two-point correlations is therefore not established as a property of the microstructure representation alone. Please repeat the evaluation with leave-one-paper-out or paper-grouped folds and report the grouped MAPE for all descriptor/model combinations; if the gap shrinks or disappears, the central hypothesis is not supported by this case study.
  2. [Section 3.2, data curation] The 149 Vickers hardness values are treated as commensurable even though they come from different papers, likely with different Vickers loads, preparation routes, and imaging conditions, and no evidence is given that the SEM image depicts the same region that was indented. This is not a minor detail: if label noise is paper-specific, a model can learn it from image statistics, which would inflate all reported accuracies and bias the descriptor comparison. Please report the test conditions available in the source papers, restrict the analysis to a commensurable subset if necessary, or add a sensitivity analysis that includes per-paper fixed effects.
  3. [Section 3.3, Figure 9] The reported sensitivity to a single fold—the SVR standard deviation drops from 12.93% to 5.42% when the fold containing the Ref. [51] alloy is removed—demonstrates that a single source can dominate a fold under the current random cross-validation scheme. The paper interprets this as a composition outlier, but it equally illustrates the source-leakage risk identified above, because the Ref. [51] sample is also a distinct image source. The grouped cross-validation requested above should be accompanied by a per-source error breakdown so that the influence of individual papers on the reported means and standard deviations can be assessed.
minor comments (4)
  1. [Sections 2, 3.1, 3.2] The required input sizes for SAM and CLIP are stated inconsistently: Section 2 and Section 3.1 imply SAM uses 224 x 224 and CLIP uses 1024 x 1024, while Section 3.2 says CLIP uses 224 x 224 and SAM uses 1024 x 1024. Please correct the earlier statements to match the actual model requirements.
  2. [Section 3.1] The statement that 42 is the multiple of 14 closest to 51 is inaccurate, since 56 is closer; the intended criterion appears to be the largest multiple of 14 not exceeding 51. Please rephrase to avoid confusion.
  3. [Figure 7] The label 'DinoV2 input' should be 'DINOv2 input' for consistency with the rest of the paper.
  4. [Section 3.2] In 'We converted the remaining data (28 values) in GPa, to the consistent units', the comma after 'GPa' is stray; also consider using 'values' instead of 'data' for subject-verb agreement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ViT features are external frozen models, the regressions are fit to independent labels, and the comparisons are honest external benchmarks.

full rationale

The paper's derivation chain is: take pre-trained ViTs with fixed weights (CLIP, DINOv2, SAM), extract features from microstructure images, reduce dimensionality with PCA, and train lightweight regressors (LR, PR, SVR) on property labels. None of these steps defines the target property in terms of the model output or fits a parameter that is later renamed as a prediction. The pre-trained ViTs are external, frozen models trained on general images, and the hardness and modulus labels come from independent experiments and simulations. The benchmark representations (two-point correlations and the MicroNet domain CNN) are also external or standard, and the comparison is a straightforward evaluation. The only tuned hyperparameter is the number of PCA components, selected on validation data via nested cross-validation; this is ordinary model selection, not circularity. The paper's self-citations (e.g., Ref. [17] for two-point correlations and Ref. [3] for related framework) support the benchmark methodology and prior context, but are not load-bearing for the central claim that pre-trained ViT features are useful for microstructure-property modeling. The acknowledged limitations about experimental data pooled from multiple papers are concerns about label commensurability and robustness, not circularity. Therefore no step in the paper reduces, by construction or self-citation, to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities. Its load-bearing assumptions are domain-level: transferability of natural-image features to microstructures, representativeness of 2D sections, commensurability of literature hardness data, and linearity of PCA. Free parameters are the usual ML hyperparameters, with the number of PCA components being the most consequential.

free parameters (4)
  • Number of principal components = 24 (two-point, case 1), 39 (SAM, case 1), 34 (DINOv2 L14, case 2), 49 (DINOv2 L14 + composition, case 2)
    Selected by grid search on validation set (case 1) or nested 10-fold CV (case 2) for each feature type and model; the final value is fit to the data and affects the reported accuracy.
  • SVR hyperparameters = Not reported
    Support vector regression in case study 2 uses C, epsilon, and kernel parameters chosen by grid search, but exact values are omitted, so the fitted model is underspecified.
  • Polynomial regression degree = 2
    Second-order polynomial chosen for PR models; a model choice not derived from theory.
  • Image preprocessing parameters = Crop sizes and resampling: e.g., 42x42 for DINOv2, 224x224 for CLIP, 1024x1024 for SAM; bilinear interpolation
    Hand-set choices that determine what the ViT sees and thus affect the feature vectors; no principled selection is given for these values.
assumptions (5)
  • domain assumption Pre-trained ViT features trained on natural images transfer to microstructure images and encode property-relevant information
    This is the central hypothesis; the paper tests it empirically rather than deriving it. It is invoked throughout the method.
  • domain assumption Three orthogonal 2D sections are sufficient to represent the 3D microstructure for predicting effective Young's modulus
    Used in case study 1, following prior work [36]; if false, the property predictions are limited by the input representation.
  • domain assumption Hardness values from 19 papers are directly commensurable and correspond to the specific SEM images used
    Case study 2 pools literature data without normalizing for indentation load, sample preparation, or imaging location; this underpins the target variable.
  • domain assumption PCA preserves the microstructure-property-relevant structure of the ViT features
    The reduced-order features are linear projections; the paper does not test whether nonlinear dimensionality reduction would change results.
  • standard math The standard implementations of PCA, SVR, and cross-validation are correct
    The paper relies on standard libraries; no formal verification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine learning of microstructure--property relationships in materials leveraging microstructure representation from foundational vision transformers." pith.science (2026). https://pith.science/paper/IUDIX6BN

@misc{pith2026250118637,
  author       = {Pith},
  title        = {Pith review of: Machine learning of microstructure--property relationships in materials leveraging microstructure representation from foundational vision transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUDIX6BN}},
  note         = {Machine review of arXiv:2501.18637}
}
read the original abstract

Machine learning of microstructure--property relationships from data is an emerging approach in computational materials science. Most existing machine learning efforts focus on the development of task-specific models for each microstructure--property relationship. We propose utilizing pre-trained foundational vision transformers for the extraction of task-agnostic microstructure features and subsequent light-weight machine learning of a microstructure-dependent property. We demonstrate our approach with pre-trained state-of-the-art vision transformers (CLIP, DINOv2, SAM) in two case studies on machine-learning: (i) elastic modulus of two-phase microstructures based on simulations data; and (ii) Vicker's hardness of Ni-base and Co-base superalloys based on experimental data published in literature. Our results show the potential of foundational vision transformers for robust microstructure representation and efficient machine learning of microstructure--property relationships without the need for expensive task-specific training or fine-tuning of bespoke deep learning models.

Figures

Figures reproduced from arXiv: 2501.18637 by the authors.

Figure 1
Figure 1. ViTs and their variants tested in this study with [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Machine learning of effective Young’s modulus ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Accuracy of Young’s modulus predictions for the test set shown in terms of MAPE for (a) linear models and (b) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Prediction of Young’s modulus shown as parity plots for a set of 590 unseen microstructures by second-order [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: PCA of microstructure features representing 5900 2D sections of two-phase microstructures: (a) low-dimensional [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Machine learning of Vicker’s hardness (HV) of superalloys using ViT-based microstructure description of grayscale [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Pre-processing of experimental microstructure images for input to the three ViTs exemplified by one of the mi [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Machine learning of Vicker’s hardness based on microstructure features shown in a (a) mean and standard deviation [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Machine learning of Vicker’s hardness based on microstructure and composition shown in a (a) mean and standard [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 47 canonical work pages

  1. [51]

    S. Chen, Q. Liu, J. Chen, L. Zhang, T. Cui, X. Sun, Microstructure and tribological performance of novel ni-based alloy cladding with excellent high tempera- ture wear resistance and self-lubrication performance, Surface and Coatings Technology 494 (2024) 131395

    work page 2024

  2. [1]

    Khatavkar, S

    N. Khatavkar, S. Swetlana, A. K. Singh, Accelerated prediction of vickers hardness of co-and ni-based super- alloys from microstructure and composition using ad- vanced image processing techniques and machine learn- ing, Acta Materialia 196 (2020) 295–303

    work page 2020

  3. [2]

    Mahdavi, M

    M. Mahdavi, M. Standish, A. Iskakov, H. Garmestani, S. R. Kalidindi, Reduced-order models correlating Ti beta 21S microstructures and Vickers hardness mea- surements, Engineering Science & Technology (2023) 69–79

    work page 2023

  4. [3]

    M. I. Latypov, L. S. Toth, S. R. Kalidindi, Materi- als knowledge system for nonlinear composites, Com- puter Methods in Applied Mechanics and Engineering 346 (2019) 180–196

    work page 2019

  5. [4]

    Ibragimova, A

    O. Ibragimova, A. Brahme, W. Muhammad, D. Con- nolly, J. L´ evesque, K. Inal, A convolutional neural net- work based crystal plasticity finite element framework to predict localised deformation in metals, International Journal of Plasticity 157 (2022) 103374

    work page 2022

  6. [5]

    G. Hu, M. I. Latypov, AnisoGNN: Graph neural net- works generalizing to anisotropic properties of poly- crystals, Computational Materials Science 243 (2024) 113121

    work page 2024

  7. [6]

    Sim, M.-G

    G.-J. Sim, M.-G. Lee, M. I. Latypov, Fip-gnn: Graph neural networks for scalable prediction of grain-level fatigue indicator parameters, Scripta Materialia 255 (2025) 116407. Postprint of Whitman & Latypov, Acta Mater (2025) 121217

    work page 2025

  8. [7]

    D. C. Pagan, C. R. Pash, A. R. Benson, M. P. Kase- mer, Graph neural network modeling of grain-scale anisotropic elastic behavior using simulated and mea- sured microscale data, npj Computational Materials 8 (1) (2022) 259

    work page 2022

Show all 57 references
  1. [8]

    S. R. Kalidindi, Hierarchical materials informatics: novel analytics for materials data, Elsevier, 2015

    2015

  2. [9]

    Hall, Variation of hardness of metals with grain size, Nature 173 (4411) (1954) 948–949

    E. Hall, Variation of hardness of metals with grain size, Nature 173 (4411) (1954) 948–949

    1954

  3. [10]

    Petch, Xvi

    N. Petch, Xvi. the ductile fracture of polycrystalline α- iron, Philosophical Magazine 1 (2) (1956) 186–190

    1956

  4. [11]

    Y. Jiao, F. Stillinger, S. Torquato, Modeling heteroge- neous materials via two-point correlation functions: Ba- sic principles, Physical review E 76 (3) (2007) 031110

    2007

  5. [12]

    B. L. Adams, T. Olson, The mesostructure–properties linkage in polycrystals, Progress in Materials Science 43 (1) (1998) 1–87. doi:10.1016/S0079-6425(98) 00002-4

    1998 doi

  6. [13]

    B. Lu, S. Torquato, Lineal-path function for random heterogeneous materials, Physical Review A 45 (2) (1992) 922

    1992

  7. [14]

    Torquato, B

    S. Torquato, B. Lu, Chord-length distribution function for two-phase random media, Physical Review E 47 (4) (1993) 2950

    1993

  8. [15]

    M. I. Latypov, M. K¨ uhbach, I. J. Beyerlein, J.-C. Stinville, L. S. Toth, T. M. Pollock, S. R. Kalidindi, Application of chord length distributions and principal component analysis for quantification and representa- tion of diverse polycrystalline microstructures, Materi- als...

    2018

  9. [16]

    S. E. Whitman, M. I. Latypov, SR-CLD: spatially- resolved chord length distributions for statistical de- scription, visualization, and alignment of non-uniform microstructures, arXiv preprint arXiv:2409.03729 (2024)

    2024 arXiv

  10. [17]

    M. I. Latypov, S. R. Kalidindi, Data-driven reduced or- der models for effective yield strength and partitioning of strain in multiphase materials, Journal of Computa- tional Physics 346 (2017) 242–261

    2017

  11. [18]

    H. Xu, R. Liu, A. Choudhary, W. Chen, A machine learning-based design representation method for design- ing heterogeneous microstructures, Journal of Mechan- ical Design 137 (5) (2015) 051403

    2015

  12. [19]

    Cecen, H

    A. Cecen, H. Dai, Y. C. Yabansu, S. R. Kalidindi, L. Song, Material structure-property linkages using three-dimensional convolutional neural networks, Acta Materialia 146 (2018) 76–84

    2018

  13. [20]

    Z. Yang, Y. C. Yabansu, R. Al-Bahrani, W.-K. Liao, A. N. Choudhary, S. R. Kalidindi, A. Agrawal, Deep learning approaches for mining structure-property link- ages in high contrast composites from simulation datasets, Computational Materials Science 151 (2018) 278–287

    2018

  14. [21]

    Pokharel, A

    R. Pokharel, A. Pandey, A. Scheinker, Physics-informed data-driven surrogate modeling for full-field 3d mi- crostructure and micromechanical field evolution of polycrystalline materials, JOM 73 (11) (2021) 3371– 3382

    2021

  15. [22]

    K. He, H. Fan, Y. Wu, S. Xie, R. Girshick, Momentum contrast for unsupervised visual representation learn- ing, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    2020

  16. [23]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Ka- plan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learn- ers, Advances in neural information processing systems 33 (2020) 1877–1901

    2020

  17. [24]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, At- tention is all you need, Advances in neural information processing systems 30 (2017)

    2017

  18. [25]

    Caron, P

    M. Caron, P. Bojanowski, A. Joulin, M. Douze, Deep clustering for unsupervised learning of visual features, in: Proceedings of the European conference on com- puter vision (ECCV), 2018, pp. 132–149

    2018

  19. [26]

    K. He, X. Chen, S. Xie, Y. Li, P. Doll´ ar, R. Girshick, Masked autoencoders are scalable vision learners, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16000–16009

    2022

  20. [27]

    T. Chen, S. Kornblith, M. Norouzi, G. Hinton, A simple framework for contrastive learning of visual representa- tions, in: International conference on machine learning, PMLR, 2020, pp. 1597–1607

    2020

  21. [28]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: International conference on machine learning, PMLR, 2021, pp. 8748–8763

    2021

  22. [29]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.- Y. Lo, et al., Segment anything, in: Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, 2023, pp. 4015–4026

    2023

  23. [30]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al., DINOv2: Learning ro- bust visual features without supervision, arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  24. [31]

    Dosovitskiy, An image is worth 16x16 words: Trans- formers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    A. Dosovitskiy, An image is worth 16x16 words: Trans- formers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  25. [32]

    McInnes, J

    L. McInnes, J. Healy, J. Melville, Umap: Uniform mani- fold approximation and projection for dimension reduc- tion, arXiv preprint arXiv:1802.03426 (2018)

    2018 arXiv

  26. [33]

    Van der Maaten, G

    L. Van der Maaten, G. Hinton, Visualizing data using t- sne., Journal of machine learning research 9 (11) (2008)

    2008

  27. [34]

    S. R. Niezgoda, A. K. Kanjarla, S. R. Kalidindi, Novel microstructure quantification framework for databas- ing, visualization, and analysis of microstructure data, Integrating Materials and Manufacturing Innovation 2 (2013) 54–80

    2013

  28. [35]

    Stuckner, B

    J. Stuckner, B. Harder, T. M. Smith, Microstructure segmentation with deep learning encoders pre-trained on a large microscopy dataset, npj Computational Ma- terials 8 (1) (2022) 200

    2022

  29. [36]

    G. Hu, M. I. Latypov, Learning from 2D: machine learn- ing of 3D effective properties of heterogeneous materials based on 2D microstructure sections, Frontiers in Met- als and Alloys 1 (2022) 1100571

    2022

  30. [37]

    M. P. Echlin, A. Mottura, C. J. Torbet, T. M. Pol- lock, A new tribeam system for three-dimensional mul- timodal materials analysis, Review of Scientific Instru- ments 83 (2) (2012) 023701

    2012

  31. [38]

    Pokharel, J

    R. Pokharel, J. Lind, S. F. Li, P. Kenesei, R. A. Leben- sohn, R. M. Suter, A. D. Rollett, In-situ observation of bulk 3d grain evolution during plastic deformation in polycrystalline cu, International Journal of Plasticity 67 (2015) 217–234

    2015

  32. [39]

    M. D. Uchic, L. Holzer, B. J. Inkson, E. L. Principe, Postprint of Whitman & Latypov, Acta Mater (2025) 121217 P. Munroe, Three-dimensional microstructural charac- terization using focused ion beam tomography, MRS bulletin 32 (5) (2007) 408–416

    2025

  33. [40]

    Cecen, S

    A. Cecen, S. R. Kalidindi, Matlab spatial correlation toolbox: Release 3.1, Integrating Materials and Manu- facturing Innovation 5 (1) (2015) 1–15. doi:10.5281/ zenodo.31329. URL https://doi.org/10.5281/zenodo.31329

    2015 doi

  34. [41]

    H. Peng, Y. Shi, S. Gong, H. Guo, B. Chen, Microstruc- ture, mechanical properties and cracking behaviour in a γ′-precipitation strengthened nickel-base superalloy fabricated by electron beam melting, Materials & De- sign 159 (2018) 155–169

    2018

  35. [42]

    Picasso, A

    A. Picasso, A. Somoza, A. Tolley, Nucleation, growth and coarsening of γ′-precipitates in a ni–cr–al-based commercial superalloy during artificial aging, Journal of alloys and compounds 479 (1-2) (2009) 129–133

    2009

  36. [43]

    Zhang, H

    Y. Zhang, H. Fu, X. Zhou, Y. Zhang, J. Xie, Effects of aluminum and molybdenum content on the microstruc- ture and properties of multi-componentγ′-strengthened cobalt-base superalloys, Materials Science and Engi- neering: A 737 (2018) 265–273

    2018

  37. [44]

    E92-17, Standard test methods for vickers hardness and knoop hardness of metallic materials (2017)

    A. E92-17, Standard test methods for vickers hardness and knoop hardness of metallic materials (2017)

    2017

  38. [45]

    Paszke, S

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, A. Lerer, Automatic differentiation in pytorch (2017)

    2017

  39. [46]

    Bradski, The OpenCV Library, Dr

    G. Bradski, The OpenCV Library, Dr. Dobb’s Journal of Software Tools (2000)

    2000

  40. [47]

    Cecen, T

    A. Cecen, T. Fast, S. R. Kalidindi, Versatile algorithms for the computation of 2-point spatial correlations in quantifying material structure, Integrating Materials and Manufacturing Innovation 5 (1) (2016) 1–15

    2016

  41. [48]

    A. J. Smola, B. Sch¨ olkopf, A tutorial on support vector regression, Statistics and computing 14 (2004) 199–222

    2004

  42. [49]

    Mesut, A

    B. Mesut, A. Ba¸ skor, N. B. Aksu, Role of artificial in- telligence in quality profiling and optimization of drug products, in: A Handbook of Artificial Intelligence in Drug Delivery, Elsevier, 2023, pp. 35–54

    2023

  43. [50]

    James, D

    G. James, D. Witten, T. Hastie, R. Tibshirani, J. Tay- lor, An introduction to statistical learning: With appli- cations in python, Springer Nature, 2023

    2023

  44. [52]

    Torquato, et al., Random heterogeneous materials: microstructure and macroscopic properties, Vol

    S. Torquato, et al., Random heterogeneous materials: microstructure and macroscopic properties, Vol. 16, Springer, 2002

    2002

  45. [53]

    Gupta, A

    A. Gupta, A. Cecen, S. Goyal, A. K. Singh, S. R. Ka- lidindi, Structure–property linkages using a data science approach: application to a non-metallic inclusion/steel composite system, Acta Materialia 91 (2015) 239–254

    2015

  46. [54]

    Bales, T

    B. Bales, T. Pollock, L. Petzold, Segmentation-free im- age processing and analysis of precipitate shapes in 2d and 3d, Modelling and Simulation in Materials Science and Engineering 25 (4) (2017) 045009

    2017

  47. [55]

    S. E. Whitman, G. Hu, H. C. Taylor, R. B. Wicker, M. I. Latypov, Automated segmentation and chord length distribution of melt pools in complex 3d printed metal artifacts, Integrating Materials and Manufacturing In- novation 13 (1) (2024) 229–243

    2024

  48. [56]

    Y. Xu, H. Weng, X. Ju, H. Ruan, J. Chen, C. Nan, J. Guo, L. Liang, A method for predicting mechani- cal properties of composite microstructure with reduced dataset based on transfer learning, Composite Struc- tures 275 (2021) 114444

    2021

  49. [57]

    Devoto, F

    A. Devoto, F. Alvetreti, J. Pomponi, P. Di Lorenzo, P. Minervini, S. Scardapane, Adaptive layer selec- tion for efficient vision transformer fine-tuning, arXiv preprint arXiv:2408.08670 (2024)

    2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.