Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Physics-Informed Neural Networks for Sparse Strain-Field Reconstruction in 4D-STEM

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a physics-informed neural network can reconstruct a continuous 2D strain tensor from 10% of the probe positions of a 4D-STEM scan, recovering the large-scale strain morphology while cutting nominal electron dose about

desk verdict Honest, reproducible methods paper with a real empirical finding, but every headline number rests on one random mask per sampling fraction, so the quantitative claims need error bars before they can be trusted. read the letter →

arxiv 2608.01601 v1 pith:I7V5MBXQ submitted 2026-08-03 cond-mat.mtrl-sci physics.ins-det

Roberto dos Reis , Gabriel T. dos Santos , Yukun Liu , Xiaobing Hu , Vinayak P. Dravid This is my paper · ORCID
classification cond-mat.mtrl-sciphysics.ins-det
keywords 4D-STEMstrainmappingphysics-informedneuralnetworkssparsereconstructionautomaticdifferentiationSaint-VenantcompatibilityBayesianuncertaintydosereduction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a physics-informed neural network can reconstruct a full 2D strain tensor from 10% of the probe positions of a conventional 4D-STEM scan, cutting the nominal electron dose about tenfold while reproducing the large-scale strain-band morphology. On an experimental 72,000-pixel strain map of a domain-structured thermoelectric material, the method reaches R² of 0.80 on the principal strain component at 10% sampling and saturates near 0.86 by 25%. At 10% sampling it reduces mean absolute error by roughly 26% over compressed sensing and 22% over Gaussian-process regression on the same sampling masks. The key mechanism is embedding two continuum-mechanics constraints, elastic equilibrium and Saint-Venant compatibility, into the network's training loss through automatic differentiation. The paper also shows that the physics prior improves accuracy only in the extreme-sparsity regime, always improves physical self-consistency, and that Bayesian variants give per-pixel uncertainty maps correlated with reconstruction error.

What carries the argument

The load-bearing element is the composite physics-plus-data loss of Eq. (10). Elastic equilibrium residuals (from ∇·σ = 0 under isotropic plane stress) and the 2D Saint-Venant residual ∂²y εxx + ∂²x εyy − 2∂x∂y εxy = 0 are evaluated by second-order automatic differentiation of the network at fresh collocation points. The named architecture is SIREN—a sine-activated implicit neural representation whose non-vanishing second derivatives make it compatible with these second-order priors. The remaining components—frozen exponential-moving-average normalization (freezing after a 200-step warmup prevents the physics term from driving the field to the trivial uniform solution), an exponential physic

What would settle it

Repeat the 10% sampling benchmark with, say, 20 independently drawn random masks on the same strain map and compute the spread of R², MAE, and the differences between the physics-informed network, compressed sensing, and Gaussian-process regression. If the spread overlaps the claimed 26% and 22% margins, the comparative claims are not settled. A second check: re-estimate the per-channel standardization statistics from the retained training pixels only, retrain, and see whether R² at 10% drops materially below 0.80; a large drop would show the reported edge partly reflects leakage from the dens

Watch

Extended reading notes

Core claim

Central claim: a physics-informed neural network reconstructs a continuous strain tensor from sparse 4D-STEM measurements, recovering chevron strain bands from 10% of probe positions with R² ≈ 0.80 on εxx at that fraction. The network is an implicit coordinate field (SIREN) outputting (εxx, εyy, εxy, θ); the loss is a masked data term plus residuals of elastic equilibrium and Saint-Venant compatibility, evaluated by automatic differentiation. Five architectural choices—implicit representation, sine activations, frozen residual normalization, an exponential physics-weight ramp, and residual-based adaptive collocation—are argued necessary for the priors to train. The paper also claims the phys

Load-bearing premise

The headline numbers rest on the assumption that a single random sampling mask per sampling fraction is representative; the paper itself notes that repeated mask draws would be needed to attach error bars to its Table 1.

Editorial extensions

If this is right

  • Experimental 4D-STEM strain maps could be acquired at roughly tenfold lower nominal electron dose by retaining 10% of probe positions, widening the range of beam-sensitive specimens accessible to quantitative strain mapping.
  • With a constitutive model matched to the specimen, the framework is claimed to transfer to semiconductors, oxides, 2D materials, and battery and thermoelectric phases.
  • Physics-consistency residuals are improved at every sampling fraction, so downstream derivative quantities such as rotation and shear gradients are more reliable than those from data-only or classical sparse reconstructions.
  • The Bayesian uncertainty maps provide a target for adaptive second-pass acquisition, concentrating new probes on the regions where the first pass had highest epistemic uncertainty.
  • The ablation supports a practical protocol: enable physics in proportion to sparsity, and treat a growing gap between physics-informed and data-only accuracy as evidence of unmodeled physics, such as domain eigenstrain.

Reading between the lines

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

  • If the single-mask caveat is resolved favourably, the natural next test is a fully prospective workflow in which per-channel standardization statistics are estimated only from the retained measurements; the paper concedes the current retrospective benchmark leaks information from the dense reference map.
  • The finding that the homogeneous equilibrium prior becomes a bias when data are abundant suggests that an eigenstrain-augmented or per-pixel learnable constitutive closure could make the physics term beneficial across the full sampling range rather than only at extreme sparsity.
  • The same second-order PDE-embedding recipe may extend to other 4D-STEM modalities, such as differential phase contrast or ptychography, by replacing the elasticity residuals with the appropriate forward-model residuals.
  • A practical falsifiable roadmap: apply the trained protocol to a fresh beam-sensitive specimen and compare its sparse reconstructions against a full-dose map of the same region, checking whether the 10% uncertainty maps identify where a targeted second pass yields the largest error reduction.
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 / 5 minor

Summary. The paper presents a physics-informed neural network (PINN) architecture for reconstructing 2D strain fields from sparsely sampled 4D-STEM data. The architecture embeds elastic equilibrium and Saint-Venant compatibility as soft constraints in the training loss, using a coordinate-based SIREN backbone with sine activations, frozen residual-scale normalization, an exponential physics-weight ramp, adaptive collocation, and two Bayesian variants (Monte Carlo dropout and mean-field variational inference) for per-pixel uncertainty. The method is demonstrated on a single experimental 180×400-pixel strain map of PbGeSnSe1.5Te1.5. Across sampling fractions of 1–75% (720–54,000 probe positions), the paper reports R²(ε_xx) rising from 0.44 at 1% to 0.80 at 10% and saturating near 0.86, MAE reductions of ~26% versus compressed sensing and ~22% versus Gaussian-process regression at 10% sampling, and an ablation showing that the physics prior improves accuracy only at extreme sparsity while always improving the imposed residual-based self-consistency metrics. The central claims are explicitly conditional on one dataset and one random mask per sampling fraction, and the paper discloses several secondary caveats.

Significance. If the quantitative claims are robust, the paper offers a practical dose-reduction protocol for 4D-STEM strain mapping, a topic of genuine interest for beam-sensitive materials. The architectural choices are well motivated by the structure of the second-order PDE constraints, and the paper is unusually honest: it reports the regime where the physics prior hurts accuracy, provides a reproducible notebook and scripts, and compares against external ground truth (the dense experimental map) rather than only fitting artifacts. The main weakness is that the headline numbers—R²=0.80 at 10%, the 26%/22% comparative margins, and the ablation percentages—are point estimates from one random mask per sampling fraction, with no error bars. The GP baseline is also capped at 1,500 training points, which disadvantages it in the very comparison used for the headline margin. These issues are fixable within the manuscript's scope, but they are load-bearing for the central quantitative claims.

major comments (4)
  1. [§3.3, Table 1, §4.4] All headline quantitative claims (R²=0.80 at 10%, MAE reductions of ~26% vs CS and ~22% vs GP, ablation percentages, saturation plateau) are point estimates from a single random sampling mask per fraction. The paper itself states in §4.4 that 'repeated mask realizations would attach error bars to Table 1.' Because the target field contains extended chevron strain bands, mask-to-mask variation at 1–10% sampling may be comparable to the reported effect sizes, and the comparative and ablation claims are not yet supported. Please run multiple mask realizations per sampling fraction and report means with confidence intervals for Table 1 and Tables S2–S4.
  2. [§3.6, Eq. (11), Table S4] The claim that the physics prior 'always improves physical self-consistency' is an evaluation of the training objective itself. The PINN explicitly minimizes Requil and Rcompat through the composite loss, while the data-only SIREN is never trained to minimize those residuals. Lower residuals for the PINN are therefore expected by construction and do not constitute independent evidence of physical consistency. Please either reframe these metrics as training-objective diagnostics or evaluate physical consistency with quantities not directly optimized, e.g., displacement compatibility obtained by integrating the reconstructed strain, or equilibrium residuals evaluated in physical units without per-component standardization.
  3. [§4.4, Table S2] The GP baseline is capped at 1,500 training points, while the 10% sampling condition supplies 7,200 retained probes. Thus the headline 22% MAE reduction relative to GP compares against a GP that saw at most 1,500 of the available points. The text limits the caveat to 'GP comparisons at ≥25% sampling understate its performance,' but the same cap binds at 10%. Please acknowledge this for the 10% comparison, or use a scalable GP approximation that uses the full retained set, or match the number of training points across methods.
  4. [§3.2–3.3, §4.4] Per-component standardization statistics are estimated from the full dense reference map and used to normalize both data and physics residuals. This is a mild information leak relative to a strictly prospective sparse acquisition, as the paper concedes. Because the reported R², MAE, and ablation percentages are obtained under this leak, they do not directly establish the prospective protocol. Please quantify the sensitivity by recomputing standardization statistics from retained measurements only (or from training splits) and report the change in the headline metrics.
minor comments (5)
  1. [Abstract, §3.5, Table 1] The abstract states R² 'saturates near 0.86 by 25%,' but Table 1 reports R²(ε_xx)=0.84 at 25%, 0.85 at 50%, and 0.86 at 75%. Please adjust the wording to agree with the table.
  2. [§3.3, Figure 3] The text says 'Fig. 3(b) shows loss trajectories and data/physics breakdown at 50% sampling,' but the decomposition is in panel (c). Please correct the cross-reference.
  3. [§2.5] The early-stopping validation split is not described. Please specify how validation points are chosen (e.g., held-out fraction of the sampled positions) and whether the validation set is excluded from the data term.
  4. [§3.7] The uncertainty-error correlations (ρ≈0.32 for MFVI, ρ≈0.27 for MCD) are point estimates from a single mask. Given the single-mask design, report intervals or at least note the single-run nature in the caption.
  5. [§2.4] The hyperparameters λ_data, α_eq, α_co, λ_max_phys, τ, ω0, and the SIREN width/depth are given in prose. A consolidated hyperparameter table would improve reproducibility.

Circularity Check

2 steps flagged · score 2.0 of 10

Central reconstruction accuracy is benchmarked against external ground truth; only the physics-consistency diagnostic and the disclosed dense-map standardization are near-tautological.

  1. self definitional [Section 3.6 'Physics-consistency diagnostics' and Eq. (10)-(11)]
    "An ablation in which the same network is trained with λphys = 0 (a data-only SIREN baseline of identical architecture and schedule) produces residuals that are 1.5–1.8× (equilibrium) and 2.2–3.9× (compatibility) larger than the PINN’s at every sampling fraction."

    The PINN's composite loss Ltotal (Eq. 10) explicitly minimizes Lequil and Lcompat, and the reported consistency metrics Requil/Rcompat (Eq. 11) are the square roots of exactly those residuals. The data-only baseline has λphys=0, so it never optimizes them. Therefore the claim that the physics term 'purchases a consistent improvement in self-consistency' is a restatement of the training objective, not an empirical discovery. This is a secondary diagnostic, not the paper's central accuracy claim.

  2. fitted input called prediction [Section 4.4, third caveat]
    "Third, per-component standardization statistics were estimated from the dense reference map for this retrospective benchmark, constituting a mild information leak relative to a strictly prospective sparse acquisition."

    The data term Ldata is defined on per-component standardized fields. The mean and standard deviation used for that standardization are computed from the full dense map, so every 'sparse' reconstruction inherits global affine statistics of the target it is later scored against. This is an admitted leak of target information into the training pipeline. It is only a mild, scalar leak and does not determine the spatial chevron structure, so it does not invalidate the central comparison, but it does blur the line between sparse prediction and partially supervised reconstruction.

full rationale

The paper's headline results—R2 values, MAE trends across sampling fractions, and the CS/GP comparisons—are all evaluated against the full dense experimental strain map, which is external to the network and not optimized by the loss except for the disclosed per-component standardization statistics. The five architectural requirements are derived from the PDE priors and instantiated with externally established components (SIREN, RAR, NTK-based ramp); the [Liu et al., 2024] citation supplies the test dataset and is not load-bearing for the method's validity. The main circular element is the physics-consistency ablation: 'PINN lowers equilibrium/compatibility residuals' is essentially guaranteed because those residuals are part of the PINN's training loss and absent from the λphys=0 baseline. This is a self-definitional consistency check, not an independent prediction, but it is not the paper's central claim. The paper itself flags the dense-map standardization leak and, in Section 4.4, the use of one dataset and one random mask per sampling fraction; the latter is a missing-support/error-bar issue, not a circularity. Overall, the core reconstruction derivation is self-contained against an external benchmark, so the circularity score is low.

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

No new physical entities are introduced. The load-bearing assumptions are the approximate constitutive model (isotropic plane-stress elasticity with a hand-picked Poisson ratio), the appropriateness of the raw noisy dense map as ground truth, the representational sufficiency of the fixed SIREN architecture, and the evaluation protocol's assumptions (single mask, dense-map-derived standardization). The free parameters are hyperparameters and statistics fitted to the test dataset, not universal constants.

free parameters (5)
  • Per-component standardization statistics (mean/std of εxx, εyy, εxy, θ) = Estimated from the full 180x400 dense reference map
    Used in the standardized data term of Eq. (10); the paper's third caveat in Section 4.4 concedes this is a mild information leak relative to a prospective sparse acquisition.
  • Poisson ratio ν = 0.27 (E = 150 GPa cancels under scale normalization)
    The equilibrium residual in Eq. (8) with Eq. (4) depends on the effective Poisson ratio; the value is chosen as representative for the specimen and is not independently measured in the paper.
  • SIREN frequency ω0 = 30
    Chosen to resolve the ~1-30 pixel features of the strain map; the paper verifies ω0=1 and ω0=60 fail, so it is a hand-tuned dataset-specific hyperparameter, not a free constant of the method.
  • Physics-weight ramp parameters (λmax_phys, τ) = 1.0, 500 epochs
    Set by hand in Eq. (10); no sensitivity study is reported, and the ramp is identified as critical for stability with second-order priors.
  • Normalization warmup length tw = 200 optimizer steps
    Frozen after warmup per Requirement 3; the choice of 200 steps is not justified by an experiment, and the claimed log-L collapse is asserted without a demonstration.
assumptions (5)
  • domain assumption Isotropic linear elasticity (Hooke's law, plane stress) governs the measured strain field.
    Used at Eqs. (3)-(4); the paper's own ablation (Table S4) shows this prior is only approximate for the ferroelastic domain structure, biasing the reconstruction at high sampling fractions.
  • domain assumption The measured strain satisfies Saint-Venant compatibility at the reconstruction scale.
    Eq. (5) is imposed as a hard regularizer; the paper acknowledges the presence of spontaneous transformation eigenstrain in the ferroelastic domains, which can violate the compatibility assumption at domain walls.
  • domain assumption The dense reference map, itself noisy (~3% pixel-scale variance), is an appropriate ground truth for R2 and MAE scoring.
    All metrics are computed against the raw full valid-pixel grid; the paper argues this understates quality, but the specific numbers depend on the noise realization of the reference and on the validity of treating it as truth.
  • domain assumption A SIREN with width 128, four hidden layers, residual skip, and ω0=30 can represent the strain field accurately enough.
    Representational capacity is assumed; the paper verifies only ω0 sensitivity and notes wider variants are learning-rate fragile, so the architecture is a fixed choice rather than a proven universal approximator for this problem class.
  • domain assumption The data noise is additive and masked MSE in Eq. (7) is an appropriate likelihood.
    The measurement model in Eq. (2) assumes additive noise; the paper notes the noise is non-stationary and generally non-Gaussian, which is then treated as MSE, potentially over-weighting outliers and masking heteroscedasticity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Informed Neural Networks for Sparse Strain-Field Reconstruction in 4D-STEM." pith.science (2026). https://pith.science/paper/I7V5MBXQ

@misc{pith2026260801601,
  author       = {Pith},
  title        = {Pith review of: Physics-Informed Neural Networks for Sparse Strain-Field Reconstruction in 4D-STEM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7V5MBXQ}},
  note         = {Machine review of arXiv:2608.01601}
}
abstract

Quantitative strain mapping using four-dimensional scanning transmission electron microscopy (4D-STEM) typically requires densely sampled scans that can damage beam-sensitive specimens. We develop a physics-informed neural network (PINN) for sparse 4D-STEM strain reconstruction that embeds elastic equilibrium and Saint-Venant compatibility in the training loss through automatic differentiation. The architecture combines a coordinate-based implicit representation, sine activations with stable second derivatives, frozen residual-scale normalization, an exponential physics-weight ramp, and residual-based adaptive collocation. We apply a sine-activated residual network to an experimental $180\times400$-pixel strain map of domain-structured PbGeSnSe$_{1.5}$Te$_{1.5}$. Across $1$-$75%$ sampling ($720$-$54{,}000$ probe positions), $R^2$ for $\varepsilon_{xx}$ reaches $0.80$ at $10%$ sampling and saturates near $0.86$ by $25%$; the chevron strain-band morphology is recovered from $10%$ of probe positions. At $10%$ sampling, the PINN reduces mean absolute error by approximately $26%$ relative to compressed sensing and $22%$ relative to Gaussian-process regression. An ablation against an equal-capacity data-only SIREN shows that the PDE prior improves accuracy at extreme sparsity and consistently improves physical self-consistency, but biases the reconstruction when data are abundant. Monte Carlo dropout and mean-field variational inference provide per-pixel epistemic uncertainty maps correlated with reconstruction error. With an appropriate constitutive model, the framework is adaptable to strain mapping across diverse material systems.

Figures

Figures reproduced from arXiv: 2608.01601 by the authors.

Figure 1
Figure 1. SIREN backbone used for the reconstruction (Section [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Experimental 4D-STEM strain map (180 × 400 pixels, all valid) used as the worked example. Panels show the three measured strain channels (εxx, εyy, εxy), the derived angle θ, the valid-pixel mask, and the per-pixel strain-magnitude map. Chevron-shaped strain bands are superposed on pixel-scale measurement noise (∼ 3% of the field variance). Scale bar is 300 nm. recorded on an aberration-corrected STEM with a pixelat… view at source ↗
Figure 3
Figure 3. (a) Representative random training-point distributions for the six sampling fractions. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: PINN reconstructions of εxx at the six sampling fractions, with retained training points overlaid in green. The chevron strain bands are recovered from 10% of the probe positions; fine structure is progressively restored as sampling increases and saturates by ∼ 25%. Pa…
Figure 5
Figure 5. Figure 5: Absolute-error heatmaps for εxx at the six sampling fractions. Error decreases rapidly up to ∼ 10% sampling and is thereafter concentrated on pixel-scale noise and the sharpest band edges, which the physics-regularized network does not attempt to reproduce. 3.7 Bayesia…
Figure 6
Figure 6. Figure 6: Reconstruction metrics versus sampling fraction. a. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Bayesian epistemic-uncertainty maps for εxx at 10% sparse data, T = 150 stochastic forward passes. Top row: ground-truth field; Monte Carlo dropout mean and standard deviation. Bottom row: training-point distribution; mean-field variational-inference mean and standard …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 51 canonical work pages

  1. [1]

    and Perdikaris, P

    Raissi, M. and Perdikaris, P. and Karniadakis, G. E. , title =. Journal of Computational Physics , volume =. 2019 , doi =

    2019

  2. [2]

    Karniadakis, G. E. and Kevrekidis, I. G. and Lu, L. and Perdikaris, P. and Wang, S. and Yang, L. , title =. Nature Reviews Physics , volume =. 2021 , doi =

    work page 2021

  3. [3]

    and Di Cola, V

    Cuomo, S. and Di Cola, V. S. and Giampaolo, F. and Rozza, G. and Raissi, M. and Piccialli, F. , title =. Journal of Scientific Computing , volume =. 2022 , doi =

    work page 2022

  4. [4]

    and Meng, X

    Lu, L. and Meng, X. and Mao, Z. and Karniadakis, G. E. , title =. SIAM Review , volume =. 2021 , doi =

    2021

  5. [5]

    and Yazdani, A

    Raissi, M. and Yazdani, A. and Karniadakis, G. E. , title =. Science , volume =. 2020 , doi =

    work page 2020

  6. [6]

    and Mao, Z

    Cai, S. and Mao, Z. and Wang, Z. and Yin, M. and Karniadakis, G. E. , title =. Acta Mechanica Sinica , volume =. 2021 , doi =

    work page 2021

  7. [7]

    and Cirac, I

    Carleo, G. and Cirac, I. and Cranmer, K. and Daudet, L. and Schuld, M. and Tishby, N. and Vogt-Maranto, L. and Zdeborov. Machine learning and the physical sciences , journal =. 2019 , doi =

    work page 2019

  8. [8]

    and Jagtap, A

    Mao, Z. and Jagtap, A. D. and Karniadakis, G. E. , title =. Computer Methods in Applied Mechanics and Engineering , volume =. 2020 , doi =

    work page 2020

Show all 55 references
  1. [9]

    and Anitescu, C

    Goswami, S. and Anitescu, C. and Chakraborty, S. and Rabczuk, T. , title =. Theoretical and Applied Fracture Mechanics , volume =. 2020 , doi =

    2020

  2. [10]

    and Martel, J

    Sitzmann, V. and Martel, J. N. P. and Bergman, A. W. and Lindell, D. B. and Wetzstein, G. , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  3. [11]

    and Srinivasan, P

    Tancik, M. and Srinivasan, P. P. and Mildenhall, B. and Fridovich-Keil, S. and Raghavan, N. and Singhal, U. and Ramamoorthi, R. and Barron, J. T. and Ng, R. , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  4. [12]

    and Yu, X

    Wang, S. and Yu, X. and Perdikaris, P. , title =. Journal of Computational Physics , volume =. 2022 , doi =

    2022

  5. [13]

    and Teng, Y

    Wang, S. and Teng, Y. and Perdikaris, P. , title =. SIAM Journal on Scientific Computing , volume =. 2021 , doi =

    2021

  6. [14]

    Krishnapriyan, A. S. and Gholami, A. and Zhe, S. and Kirby, R. M. and Mahoney, M. W. , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  7. [15]

    McClenny, L. D. and Braga-Neto, U. M. , title =. Journal of Computational Physics , volume =. 2023 , doi =

    2023

  8. [16]

    and Zhu, M

    Wu, C. and Zhu, M. and Tan, Q. and Kartha, Y. and Lu, L. , title =. Computer Methods in Applied Mechanics and Engineering , volume =. 2023 , doi =

    2023

  9. [17]

    and Zhang, X

    He, K. and Zhang, X. and Ren, S. and Sun, J. , title =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2016 , doi =

    2016

  10. [18]

    Kingma, D. P. and Ba, J. , title =. International Conference on Learning Representations (ICLR) , year =

  11. [19]

    and Gross, S

    Paszke, A. and Gross, S. and Massa, F. and Lerer, A. and Bradbury, J. and others , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  12. [20]

    and Raissi, M

    Haghighat, E. and Raissi, M. and Moure, A. and Gomez, H. and Juanes, R. , title =. Computer Methods in Applied Mechanics and Engineering , volume =. 2021 , doi =

    2021

  13. [21]

    and Wessels, H

    Henkes, A. and Wessels, H. and Mahnken, R. , title =. Computer Methods in Applied Mechanics and Engineering , volume =. 2022 , doi =

    2022

  14. [22]

    and Sun, H

    Rao, C. and Sun, H. and Liu, Y. , title =. Journal of Engineering Mechanics , volume =. 2021 , doi =

    2021

  15. [23]

    and Meng, X

    Yang, L. and Meng, X. and Karniadakis, G. E. , title =. Journal of Computational Physics , volume =. 2021 , doi =

    2021

  16. [24]

    and Ghahramani, Z

    Gal, Y. and Ghahramani, Z. , title =. International Conference on Machine Learning (ICML) , pages =

  17. [25]

    and Cornebise, J

    Blundell, C. and Cornebise, J. and Kavukcuoglu, K. and Wierstra, D. , title =. International Conference on Machine Learning (ICML) , pages =

  18. [26]

    Tartakovsky, A. M. and Marrero, C. O. and Perdikaris, P. and Tartakovsky, G. D. and Barajas-Solano, D. , title =. Water Resources Research , volume =. 2020 , doi =

    2020

  19. [27]

    , title =

    Ophus, C. , title =. Microscopy and Microanalysis , volume =. 2019 , doi =

    2019

  20. [28]

    and Holtz, M

    Padgett, E. and Holtz, M. E. and Cueva, P. and Shao, Y.-T. and Langenberg, E. and Schlom, D. G. and Muller, D. A. , title =. Ultramicroscopy , volume =. 2020 , doi =

    2020

  21. [29]

    Pekin, T. C. and Gammer, C. and Ciston, J. and Minor, A. M. and Ophus, C. , title =. Ultramicroscopy , volume =. 2017 , doi =

    2017

  22. [30]

    Scanning transmission electron microscopy strain measurement from millisecond frames of a direct electron charge coupled device , journal =

    M. Scanning transmission electron microscopy strain measurement from millisecond frames of a direct electron charge coupled device , journal =. 2012 , doi =

    2012

  23. [31]

    Hachtel, J. A. and Idrobo, J. C. and Chi, M. , title =. Advanced Structural and Chemical Imaging , volume =. 2018 , doi =

    2018

  24. [32]

    Egerton, R. F. , title =. Micron , volume =. 2019 , doi =

    2019

  25. [33]

    Savitzky, B. H. and Zeltmann, S. E. and Hughes, L. A. and Brown, H. G. and Zhao, S. and Pelz, P. M. and Pekin, T. C. and Barnard, E. S. and Donohue, J. and DaCosta, L. and others , title =. Microscopy and Microanalysis , volume =. 2021 , doi =

    2021

  26. [34]

    Ultramicroscopy , pages=

    Free, flexible and fast: Orientation mapping using the multi-core and GPU-accelerated template matching capabilities in the python-based open source 4D-STEM analysis toolbox Pyxem , author=. Ultramicroscopy , pages=. 2022 , publisher=

    2022

  27. [35]

    and Rakowski, A

    Munshi, J. and Rakowski, A. and Savitzky, B. H. and Zeltmann, S. E. and Ciston, J. and Henderson, M. and Cholia, S. and Minor, A. M. and Chan, M. K. Y. and Ophus, C. , title =. npj Computational Materials , volume =. 2022 , doi =

    2022

  28. [36]

    and Ghosh, A

    Ziatdinov, M. and Ghosh, A. and Wong, C. Y. and Kalinin, S. V. , title =. Nature Machine Intelligence , volume =. 2022 , doi =

    2022

  29. [37]

    and Liu, P

    Madsen, J. and Liu, P. and Kling, J. and Wagner, J. B. and Hansen, T. W. and Winther, O. and Schi. A deep learning approach to identify local structures in atomic-resolution transmission electron microscopy images , journal =. 2018 , doi =

    2018

  30. [38]

    and Henninen, T

    Wang, F. and Henninen, T. R. and Keller, D. and Erni, R. , title =. Applied Microscopy , volume =. 2020 , doi =

    2020

  31. [39]

    and Camps-Valls, G

    Reichstein, M. and Camps-Valls, G. and Stevens, B. and Jung, M. and Denzler, J. and Carvalhais, N. and Prabhat , title =. Nature , volume =. 2019 , doi =

    2019

  32. [40]

    and Yang, H

    Stevens, A. and Yang, H. and Hao, W. and Jones, L. and Ophus, C. and Nellist, P. D. and Browning, N. D. , title =. Applied Physics Letters , volume =. 2018 , doi =

    2018

  33. [41]

    and Stevens, A

    Kovarik, L. and Stevens, A. and Liyu, A. and Browning, N. D. , title =. Applied Physics Letters , volume =. 2016 , doi =

    2016

  34. [42]

    Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information , journal =

    Cand. Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information , journal =. 2006 , doi =

    2006

  35. [43]

    Donoho, D. L. , title =. IEEE Transactions on Information Theory , volume =. 2006 , doi =

    2006

  36. [44]

    Press, W. H. and Teukolsky, S. A. and Vetterling, W. T. and Flannery, B. P. , title =

  37. [45]

    Rasmussen, C. E. and Williams, C. K. I. , title =

  38. [46]

    Journal of the American Chemical Society , volume=

    Implications and optimization of domain structures in IV--VI high-entropy thermoelectric materials , author=. Journal of the American Chemical Society , volume=. 2024 , publisher=

    2024

  39. [47]

    Nature materials , volume=

    Towards data-driven next-generation transmission electron microscopy , author=. Nature materials , volume=. 2021 , publisher=

    2021

  40. [48]

    Acta Crystallographica Section A , volume =

    Hoppe, Walter , title =. Acta Crystallographica Section A , volume =. 1969 , doi =

    1969

  41. [49]

    Acta Crystallographica Section A , volume =

    Hoppe, Walter and Strube, Gerhard , title =. Acta Crystallographica Section A , volume =. 1969 , doi =

    1969

  42. [50]

    and McCallum, Brian C

    Nellist, Peter D. and McCallum, Brian C. and Rodenburg, John M. , title =. Nature , volume =. 1995 , doi =

    1995

  43. [51]

    and Bates, Robin H

    Rodenburg, John M. and Bates, Robin H. T. , title =. Philosophical Transactions of the Royal Society of London. Series A: Physical and Engineering Sciences , volume =. 1992 , doi =

    1992

  44. [52]

    and Yang, Hao and Jones, Lewys and Lupini, Andrew R

    Pennycook, Timothy J. and Yang, Hao and Jones, Lewys and Lupini, Andrew R. and Murfitt, Matthew F. and Nellist, Peter D. , title =. Ultramicroscopy , volume =. 2015 , doi =

    2015

  45. [53]

    and Maiden, Andrew M

    Rodenburg, John M. and Maiden, Andrew M. , title =. Springer Handbook of Microscopy , editor =. 2019 , pages =

    2019

  46. [54]

    Physical Review A , volume =

    Lubk, Axel and Zweck, Josef , title =. Physical Review A , volume =. 2015 , doi =

    2015

  47. [55]

    Scientific reports , volume=

    Imaging of built-in electric field at a pn junction by scanning transmission electron microscopy , author=. Scientific reports , volume=. 2015 , publisher=

    2015

Pith tools

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