Imaging 3D polarization dynamics via deep learning 4D-STEM

Eunha Lee; Geun Ho Gu; Haeryong Kim; Jeongil Bang; Jinho Byun; Keeyong Lee; Myoungho Jeong; Sang Ho Oh

REVIEW 4 major objections 5 minor 59 references

Imaging 3D polarization dynamics via deep learning 4D-STEM

T0 review · 4 major / 5 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 A deep learning analysis of 4D-STEM images reconstructs all three polarization components, including the one along the electron beam, with sub-picometer error, and shows ferroelectric switching proceeds by rotation.

desk verdict A genuinely new 3D polarization reconstruction pipeline with strong simulation validation, but the experimental claims rest on an unverified domain-adaptation assumption and need code/data release plus an independent check. read the letter →

arxiv 2506.06598 v1 pith:GLBVN3UB submitted 2025-06-07 cond-mat.mtrl-sci

Jinho Byun , Keeyong Lee , Myoungho Jeong , Eunha Lee , Jeongil Bang , Haeryong Kim , Geun Ho Gu , Sang Ho Oh This is my paper

classification cond-mat.mtrl-sci

keywords 4D-STEMdeeplearningpolarizationmappingferroelectricsrelaxorbariumstrontiumtitanateWassersteinautoencoderdomainadaptation

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 Wasserstein autoencoder combined with a predictor network, trained on 103,200 multislice-simulated diffraction patterns and adapted to experimental data by a domain adaptation layer, can reconstruct all three components of the polarization vector in a ferroelectric thin film from 4D-STEM data, including the component along the electron beam that conventional electron microscopy cannot access. Applied to a Ba0.5Sr0.5TiO3 thin-film capacitor under bias, the method produces 3D polarization maps with mean absolute error below 1 pm. These maps show densely packed polar nanodomains, switching by coordinated vector rotation toward <111> energy minima rather than by magnitude change, and smaller polarization response in regions of high topological density. The stated significance is that full 3D polarization mapping is necessary to correctly identify polar nanodomains and to understand relaxor-like switching, with implications for ferroelectric devices.

What carries the argument

The argument is carried by a Wasserstein autoencoder (WAE) trained jointly to reconstruct diffraction patterns and predict physical parameters from a 32-dimensional latent space, with a domain adaptation (DA) layer that transforms experimental diffraction patterns into the simulated distribution before encoding. Training data are 103,200 multislice-simulated diffraction patterns generated from molecular-dynamics BST structures spanning polarization states, tilts, and thicknesses, and the DA layer raises the overlap between experimental and simulated latent distributions from 51 percent to 91 percent. The predictor resolves parameters so cleanly that individual latent dimensions correlate with individual physical properties (L2/L10 with Py, L17/L19 with Pz), evidence that the network has disentangled the polarization signal from tilt and thickness.

What would settle it

Take an experimental 4D-STEM dataset from a BST or similar film with a deliberately engineered, known out-of-plane polarization pattern, run the domain-adapted model, and compare the reconstructed Pz against the known pattern; alternatively, feed the same experimental patterns through the model with the DA layer disabled and check whether the polarization predictions degrade while tilt and thickness predictions stay stable. A clean negative result would be a case where the DA-transformed pattern shows large intensity changes in regions where the physical polarization is known to be constant.

Watch

Extended reading notes

Core claim

The central discovery is that the electron-beam-direction component of polarization, Pz, is recoverable from 4D-STEM diffraction patterns at picometer-level accuracy, and that including it changes the physical picture of a relaxor-like ferroelectric. On simulated test data the model predicts Px, Py, Pz, thickness, and tilts with mean absolute error below 1 pm for polarization. On experimental BST data, the 3D maps match molecular dynamics statistics, reveal that apparent nonpolar regions are actually nanodomains with large Pz, and show that under ±2 V bias the polarization vectors rotate toward the nearest {111} energy minima, narrowing the misorientation distribution from about 53 degrees FWHM to roughly 30 to 35 degrees, while polarization magnitude and domain size remain essentially unchanged. The paper also claims that regions of higher topological density show smaller polarization change under bias, which it interprets as topological protection.

Load-bearing premise

The result stands or falls on the assumption that the domain adaptation layer transforms experimental diffraction patterns into the simulated distribution without touching the subtle intensity differences that encode polarization; if that transformation also cleans out or reshapes the polarization signal, the experimental maps and the switching and topology conclusions would be artifacts of the simulation manifold rather than measurements.

Editorial extensions

If this is right

If the claim holds, conventional 2D in-plane polarization maps of relaxor-like films should be reinterpreted: regions that look nonpolar may simply be dominated by Pz, and the polar slush picture rather than isolated polar nanoregions describes the zero-field state.
Polarization switching in BST under electric field should be modeled as rotation in 3D toward <111> minima, not as growth or shrinkage of domains with fixed polarization magnitude; the measured angular narrowing under bias gives a quantitative target for such models.
Topological density becomes a predictor of switching resilience: domain configurations with higher local topological density should respond less to applied fields, which is testable in other ferroelectric heterostructures.
The same WAE plus DA pipeline should be transferable to other materials and beam conditions, provided the simulation manifold covers their thickness, tilt, and polarization ranges.

Reading between the lines

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

A testable extension: applying the same framework to a sample with a known uniform out-of-plane polarization component, or to a ferroelectric with a different switching axis, would separate the method's general validity from BST-specific training.
Because the DA layer is fitted to match the simulation manifold, the reported sub-picometer accuracy is a simulation-domain accuracy; the true experimental accuracy could be bounded by cross-validating against an independent thickness measurement and known field-induced displacements, and by checking whether DA-transformed patterns preserve the small Pz-sensitive intensity derivatives.
The observed narrowing of misorientation under field resembles what a soft-mode or Goldstone-like energy landscape would predict; if confirmed, it suggests device strategies that tune domain topology rather than domain size to control switching.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The new thing here is real: a WAE + predictor + domain-adaptation layer that reconstructs all three polarization components—including Pz along the beam—from 4D-STEM. I don't see that combination in the cited ML-4D-STEM literature, and the simulation-side validation is unusually thorough. 103,200 simulated DPs, out-of-sample MAE under 1 pm, explicit intensity-derivative analysis showing Pz is distinguishable from tilt and thickness, and a fair head-to-head against HAADF picking and Friedel-pair asymmetry on a simulated gradient sample. The EELS thickness agreement is a good independent check for a strong, slowly varying parameter. They also honestly flag the low-dose limit. Credit where due: this is a serious methodological step.

Now the soft spots. The stress-test note is on target. The DA layer is trained to align latent distributions (MMD) and reconstruct DPs; nothing in the objective forces it to preserve the weak, polarization-dependent intensity asymmetries. The latent overlap rising from 51% to 91% and the nearest-simulated-DP similarity are distribution-matching metrics, not signal-preservation metrics. So the experimental Px/Py/Pz maps, and everything built on them—polar nanodomains, rotation toward <111>, topological protection—are projections onto the simulation manifold. There is no independent ground truth for experimental polarization, and the promised code/model only after acceptance blocks inspection of the learned DA weights. That is a load-bearing gap, not a cosmetic one.

That said, the paper is not sloppy. The MD simulations, the PES calculations, and the qualitative consistency of the experimental polarization maps with known BST behavior make the physical story plausible. The authors clearly know the diffraction physics; the derivative analysis and latent-disentanglement checks are exactly the right kind of sanity checks. But plausible is not the same as demonstrated, and the central experimental claim needs either a reference-sample benchmark, a synthetic experimental-like transfer test, or at minimum public release of code and trained weights so others can probe the DA transformation.

Who benefits: microscopists and ferroelectrics researchers who want a route to beam-direction polarization. It deserves serious refereeing, not desk rejection. I'd send it to review with a strong request for code/data release and an explicit demonstration—on simulated data with a known DA-like perturbation—that the DA layer preserves polarization information. If that check fails, the experimental conclusions would need to be scaled back to "consistent with simulation," which is still interesting but not the full claim.

Referee Report

4 major / 5 minor

Summary. The paper presents a deep learning framework for reconstructing all three components of the polarization vector, including the electron-beam-direction component Pz, from four-dimensional scanning transmission electron microscopy (4D-STEM) data. A Wasserstein autoencoder with an auxiliary predictor network is trained on simulated diffraction patterns generated from molecular-dynamics-derived Ba0.5Sr0.5TiO3 structures, and a domain adaptation layer is introduced to transform experimental diffraction patterns into the simulated domain. The authors report mean absolute errors below 1 pm on simulated out-of-sample data, demonstrate parameter identifiability via intensity derivatives, and apply the method to in-situ biased BST thin-film capacitors. From the experimental polarization maps, they conclude the presence of polar nanodomains consistent with the polar slush model, switching via coordinated vector rotation toward <111> energy minima rather than magnitude changes, and topological protection in regions of high topological density.

Significance. If the method is sound, it addresses a genuine and important gap: the lack of sensitivity of conventional STEM polarization mapping to the polarization component along the electron beam. The paper's simulation-based validation is extensive, including out-of-sample parity plots, intensity derivative analysis supporting independent identifiability of Pz, and a dose-dependence study that quantifies experimental feasibility. These are concrete strengths and provide a credible basis for the approach on simulated data. The potential impact of a validated 3D polarization mapping technique for relaxor-like ferroelectrics is substantial. However, the central experimental claims currently rest on an unverified assumption about the domain adaptation layer, and the lack of any experimental ground truth for the polarization vector means the reported maps and derived physical conclusions are not yet independently established.

major comments (4)

[Deep learning architecture; Model Training] The domain adaptation (DA) layer is trained solely to align the latent distributions of experimental and simulated diffraction patterns (via MMD) and to reconstruct DPs; there is no constraint in Eq. (3) that the DA transformation preserve polarization-dependent intensity variations. Because all experimental polarization maps are obtained after DA, any intensity change that the DA layer removes or imprints will directly corrupt the predicted Px, Py, and Pz. The validations shown in Extended Data Fig. 2 are parity plots on simulated data, not on DA-transformed experimental data with known labels, and Extended Data Fig. 1c,d demonstrate distribution overlap and nearest-simulated-DP similarity, which are distribution-matching metrics rather than signal-preservation metrics. The authors should add a quantitative test, for example applying the DA layer to simulated DPs with known Pz and checking that predicted Pz after transformation remains accurate, or designing a physical phantom with known polarization, or adding a cycle-consistency or physics-based invariance loss that explicitly penalizes removal of the polarization signal.
[Extended Data Fig. 2] The paper claims picometer-level accuracy (MAE < 1 pm) for polarization components, but it is unclear whether Extended Data Fig. 2 is evaluated on raw simulated test patterns or on simulated patterns after the DA layer has been applied. If the MAE is obtained on clean, untransformed simulated DPs, it is not representative of the experimental pipeline, where DA necessarily alters the input. The authors should state explicitly which data pipeline produced the parity plots and, if DA is applied, include the transformation in the test procedure.
[3D polarization mapping revealing polar nanodomains and their topologies] The experimental 3D maps, domain sizes, and topological density are derived entirely from DA-transformed experimental patterns, and no independent experimental measurement of Px, Py, or Pz is provided. The EELS thickness comparison in Extended Data Fig. 6 validates a strong, slowly varying parameter and cannot validate the weak Pz signal that the method is designed to detect. Without an external benchmark for the polarization vector, the conclusions about polar nanodomains and topological protection are not falsifiable in this manuscript. The authors should either provide such a benchmark or clearly frame these results as model-dependent predictions that require further verification.
[Model architecture; Model Training] The functional form of the DA layer is not described: it is not specified whether it is a learned per-pixel gain/offset, a convolutional filter bank, or a more complex transformation, nor what constraints (e.g., smoothness, boundedness) are imposed. This information is essential for assessing whether the DA layer could selectively discard the subtle disk-intensity asymmetries that encode Pz. Additionally, no sensitivity analysis is provided for the loss weights (λrecon, λMMD, λpred) with respect to signal preservation; the chosen weights (1, 100, 5) place heavy emphasis on distribution matching, which may further encourage removal of experimental features not present in the simulation manifold.

minor comments (5)

[Abstract; Methods (4D STEM data acquisition)] The probe convergence semi-angle is stated as 2 mrad in the abstract and main text but as 2.2 mrad in the Methods; please reconcile these values.
[Data availability] The data availability statement says the machine learning model will be available only after acceptance; for evaluation, a version should be provided as supplementary material or via an anonymized repository during review.
[Model Training (Eq. 3)] Equation (3) is labeled Ltotal, but Eq. (1) is the structure factor and Eq. (2) is the topological charge; the equation numbering appears inconsistent and should be fixed.
[Deep learning architecture; Methods (4D STEM data acquisition)] The manuscript states a 63 × 165-pixel grid in the architecture section but later reports 180 × 110 probe positions in Methods; please clarify the actual scan dimensions used.
[Potential energy surface calculation] The minimal energy path (MEP) calculation is described only as 'gradient descent starting from the local minima'; please specify the initial points, the convergence tolerance, and the number of images along the path for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the prediction pipeline is trained on independent simulations and validated on held-out simulated data; the domain-adaptation assumption is a robustness risk, not a circular reduction.

full rationale

The paper's derivation chain is not circular. Polarization values are defined by atomistic MD structures and mapped to diffraction patterns by independent multislice simulations; the neural predictor is then trained on simulated (diffraction-pattern, polarization) pairs. Experimental diffraction patterns are mapped into the simulated domain by a learned DA layer, after which the same encoder predicts Px, Py, and Pz. No experimental polarization value is fitted and then re-predicted, and no equation defines the target quantity in terms of the model output. The held-out simulated parity plots (Extended Data Fig. 2) provide an internal check that the predictor can recover the simulated mapping, and the EELS thickness comparison provides an independent external check for one parameter. The main limitation is that the DA alignment is trained only to match latent distributions, not explicitly to preserve the weak Pz-dependent signal; this is an unverified domain-shift assumption and a correctness risk, but it is not a circular reduction. The paper does not rely on load-bearing self-citations, and no result is equivalent to its inputs by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The reconstruction pipeline relies on a chain of simulation assumptions. The most fragile is the DA layer and the simulation fidelity, since there is no independent experimental measurement of polarization to calibrate or verify the outputs.

free parameters (2)

Loss weights (λrecon, λMMD, λpred) = (1, 100, 5)
Hyperparameters chosen by hand to balance reconstruction, MMD regularization, and prediction error; not derived from first principles.
Latent space dimension = 24
Dimensionality of the latent vector, chosen as a design parameter; affects disentanglement and reconstruction.

assumptions (4)

domain assumption Multislice simulation (abTEM) accurately reproduces experimental 4D-STEM diffraction intensities for the BST sample.
The model is trained on simulated DPs only. If the simulator omits or misrepresents physics (e.g., inelastic scattering, detector noise, contamination), the learned mapping may not transfer to experiment.
domain assumption Machine-learning molecular dynamics generates BST structures whose polarization states cover the full experimental parameter space.
The training labels (Px, Py, Pz) are derived from MLMD snapshots. If MD does not reproduce the true distribution of polarization states, the model cannot learn the correct inversion.
ad hoc to paper The domain adaptation layer removes experimental artifacts while preserving the polarization-dependent signal.
The DA transformation is a learned mapping from experimental to simulated-like DPs; there is no proof that it does not also alter the polarization information. This assumption is critical to trusting the experimental polarization maps.
standard math The relation between atomic displacement and diffraction intensity is invertible under the simulation settings (thickness 30-80 nm, tilt ±2 mrad).
The model assumes the mapping from parameters to DP is one-to-one; the intensity derivative analysis supports local identifiability, but global invertibility is not proven.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Imaging 3D polarization dynamics via deep learning 4D-STEM." pith.science (2026). https://pith.science/paper/GLBVN3UB

@misc{pith2026250606598,
  author       = {Pith},
  title        = {Pith review of: Imaging 3D polarization dynamics via deep learning 4D-STEM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GLBVN3UB}},
  note         = {Machine review of arXiv:2506.06598}
}

read the original abstract

Recent advances in ferroelectrics highlight the role of three-dimensional (3D) polar entities in forming topological polar textures and generating giant electromechanical responses, during polarization rotation. However, current electron microscopy methods lack the depth resolution to resolve the polarization component along the electron beam direction, which restricts full characterization. Here, we present a deep learning framework combined with four-dimensional scanning transmission electron microscopy to reconstruct 3D polarization maps in Ba0.5Sr0.5TiO3 thin-film capacitors with picometer-level accuracy under applied electric fields. Our approach enables observation of polar nanodomains consistent with the polar slush model and shows that switching occurs through coordinated vector rotation toward <111> energy minima, rather than magnitude changes. Furthermore, regions with higher topological density exhibit smaller polarization variation when the electric field changes, indicating topological protection. Our work reveals the value of 3D polarization mapping in elucidating complex nanoscale polar phenomena, with broad implications for emergent ferroelectrics.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 53 canonical work pages

[1]

Cohen, R. E. Relaxors go critical. Nature 441, 941–942 (2006)

work page 2006
[2]

Yin, J. et al. Nanoscale bubble domains with polar topologies in bulk ferroelectrics. Nat Commun 12, 3632 (2021)

work page 2021
[3]

Yadav, A. K. et al. Observation of polar vortices in oxide superlattices. Nature 530, 198–201 (2016)

2016
[4]

Bencan, A. et al. Atomic scale symmetry and polar nanoclusters in the paraelectric phase of ferroelectric materials. Nat Commun 12, 3509 (2021)

work page 2021
[5]

Pan, H. et al. Ultrahigh energy storage in superparaelectric relaxor ferroelectrics. Science 374, 100–104 (2021)

work page 2021
[6]

Kumar, A. et al. Atomic-resolution electron microscopy of nanoscale local structure in lead-based relaxor ferroelectrics. Nat. Mater. 20, 62–67 (2021)

work page 2021
[7]

Gong, F. -H. et al. Atomic mapping of periodic dipole waves in ferroelectric oxide. Science Advances 7, eabg5503 (2021)

work page 2021
[8]

& Rappe, A

Takenaka, H., Grinberg, I., Liu, S. & Rappe, A. M. Slush -like polar structures in single -crystal relaxors. Nature 546, 391–395 (2017)

work page 2017

Show all 59 references

[9]

Eremenko, M. et al. Local atomic order and hierarchical polar nanoregions in a classical relaxor ferroelectric. Nat Commun 10, 2728 (2019)

2019
[10]

Liu, H. et al. Emergence of high piezoelectricity from competing local polar order -disorder in relaxor ferroelectrics. Nat Commun 14, 1007 (2023)

2023
[11]

Das, S. et al. Observation of room-temperature polar skyrmions. Nature 568, 368–372 (2019)

2019
[12]

Kim, J. et al. Size-driven phase evolution in ultrathin relaxor films. Nat. Nanotechnol. 1–9 (2025) doi:10.1038/s41565-025-01863-x

2025 doi
[13]

Jeong, J. et al. Subsurface Oxygen Vacancy Mediated Surface Reconstruction and Depolarization of Ferroelectric BaTiO3 (001) Surface. Advanced Science 2412781
[14]

Pennington, R. S. & Koch, C. T. A three-dimensional polarization domain retrieval method from electron diffraction data. Ultramicroscopy 155, 42–48 (2015)

2015
[15]

Goossens, D. J. Diffuse Scattering from Lead -Containing Ferroelectric Perovskite Oxides. International Scholarly Research Notices 2013, 107178 (2013)

2013
[16]

& Krisch, M

Bosak, A., Chernyshov, D., Vakhrushev, S. & Krisch, M. Diffuse scattering in relaxor ferroelectrics: true three-dimensional mapping, experimental artefacts and modelling. Acta Cryst A 68, 117–123 (2012)

2012
[17]

Krogstad, M. J. et al. The relation of local order to material properties in relaxor ferroelectrics. Nature Mater 17, 718–724 (2018)

2018
[18]

Kim, J. et al. Coupled polarization and nanodomain evolution underpins large electromechanical responses in relaxors. Nat. Phys. 18, 1502–1509 (2022)

2022
[19]

Shao, Y.-T. et al. Emergent chirality in a polar meron to skyrmion phase transition. Nat Commun 14, 1355 (2023)

2023
[20]

& Paik, H

Bae, I.-T., Foran, B. & Paik, H. Four dimensional -scanning transmission electron microscopy 15 study on relationship between crystallographic orientation and spontaneous polarization in epitaxial BiFeO3. Sci Rep 14, 15513 (2024)

2024
[21]

Ludacka, U. et al. Imaging and structure analysis of ferroelectric domains, domain walls, and vortices by scanning electron diffraction. npj Comput Mater 10, 1–8 (2024)

2024
[22]

Han, L. et al. High-density switchable skyrmion -like polar nanodomains integrated on silicon. Nature 603, 63–67 (2022)

2022
[23]

Kaufmann, K. et al. Crystal symmetry determination in electron diffraction using machine learning. Science (2020) doi:10.1126/science.aay3062

2020 doi
[24]

Gleason, S. P. et al. Random Forest Prediction of Crystal Structure from Electron Diffraction Patterns Incorporating Multiple Scattering. arXiv.org https://arxiv.org/abs/2406.16310v1 (2024)

2024 arXiv
[25]

Shi, C. et al. Uncovering material deformations via machine learning combined with four - dimensional scanning transmission electron microscopy. npj Comput Mater 8, 1–9 (2022)

2022
[26]

Li, X. et al. Unsupervised Machine Learning to Distill Structural -Property Insights from 4D - STEM. Microsc Microanal 25, 12–13 (2019)

2019
[27]

Kimoto, K. et al. Unsupervised machine learning combined with 4D scanning transmission electron microscopy for bimodal nanostructural analysis. Sci Rep 14, 2901 (2024)

2024
[28]

S., Coll, C., Estradé, S., Peiró, F

Pennington, R. S., Coll, C., Estradé, S., Peiró, F. & Koch, C. T. Neural -network-based depth- resolved multiscale structural optimization using density functional theory and electron diffraction data. Phys. Rev. B 97, 024112 (2018)

2018
[29]

Zeltmann, S. E. et al. Uncovering polar vortex structures by inversion of multiple scattering with a stacked Bloch wave model. Ultramicroscopy 250, 113732 (2023)

2023
[30]

Kim, K.-H., Payne, D. A. & Zuo, J. M. Determination of fluctuations in local symmetry and measurement by convergent beam electron diffraction: applications to a relaxor -based ferroelectric crystal after thermal annealing. J Appl Cryst 46, 1331–1337 (2013)

2013
[32]

Omidvar, M. et al. Accelerated discovery of perovskite solid solutions through automated materials synthesis and characterization. Nat Commun 15, 6554 (2024)

2024
[33]

& Son, J

Ahn, Y. & Son, J. Y. Enhanced energy storage performance in Ba1 -xSrxTiO3 thin films driven by relaxor ferroelectric behavior. Ceramics International (2025) doi:10.1016/j.ceramint.2025.03.274

2025 doi
[34]

Garten, L. M. et al. Relaxor Ferroelectric Behavior in Barium Strontium Titanate. Journal of the American Ceramic Society 99, 1645–1650 (2016)

2016
[35]

Lupi, E. et al. Engineering Relaxor Behavior in (BaTiO3)/(SrTiO3) Superlattices. Advanced Materials 35, 2302012 (2023)

2023
[36]

& Deng, X

Cai, W., Fan, Y., Gao, J., Fu, C. & Deng, X. Microstructure, dielectric properties and diffuse phase transition of barium stannate titanate ceramics. J Mater Sci: Mater Electron 22, 265–272 (2011)

2011
[37]

E., Dkhil, B

Sassi, E., Maayoufi, A. E., Dkhil, B. & Jomni, F. Enhancing Structural, electrical and energy storage performance of lead -free BaxSr1 -xTiO3 ceramics. Ceramics International (2025) doi:10.1016/j.ceramint.2025.04.247

2025 doi
[38]

Xiao, B. et al. Effect of large strain on dielectric and ferroelectric properties of Ba0.5Sr0.5TiO3 thin films. Applied Physics Letters 95, 012907 (2009)

2009
[39]

Li, D., Yang, Y., Song, Y. -Z. & Hospedales, T. M. Deeper, broader and artier domain generalization. in Proceedings of the IEEE international conference on computer vision 5542– 5550 (2017)

2017
[40]

& Lempitsky, V

Ganin, Y. & Lempitsky, V. Unsupervised Domain Adaptation by Backpropagation. Preprint at https://doi.org/10.48550/arXiv.1409.7495 (2015). 16

2015 doi arXiv
[41]

& Yuan, Y

Yang, C., Guo, X., Chen, Z. & Yuan, Y. Source free domain adaptation for medical image segmentation with fourier style mining. Medical Image Analysis 79, 102457 (2022)

2022
[42]

& Wang, J

Liu, Y., Zhang, W. & Wang, J. Source -Free Domain Adaptation for Semantic Segmentation. Preprint at https://doi.org/10.48550/arXiv.2103.16372 (2021)

2021 doi arXiv
[43]

& Liu, M

Fang, Y., Yap, P.-T., Lin, W., Zhu, H. & Liu, M. Source-Free Unsupervised Domain Adaptation: A Survey. Preprint at https://doi.org/10.48550/arXiv.2301.00265 (2023)

2023 doi arXiv
[44]

& Borisevich, A

Laanait, N., Ziatdinov, M., He, Q. & Borisevich, A. Identifying local structural states in atomic imaging by computer vision. Advanced structural and chemical imaging 2, 14 (2016)

2016
[45]

& Tanaka, M

Tsuda, K., Sano, R. & Tanaka, M. Nanoscale local structures of rhombohedral symmetry in the orthorhombic and tetragonal phases of BaTiO${}_{3}$ studied by convergent -beam electron diffraction. Phys. Rev. B 86, 214106 (2012)

2012
[46]

Shao, Y. -T. & Zuo, J. -M. Nanoscale symmetry fluctuations in ferroelectric barium titanate, BaTiO3. Acta Cryst B 73, 708–714 (2017)

2017
[47]

& Høier, R

Marthinsen, K. & Høier, R. On the breakdown of Friedel’s law in electron backscattering channelling patterns. Foundations of Crystallography 44, 700–707 (1988)

1988
[48]

Bird, D. M. & Wright, A. G. Phase dependence of Kikuchi patterns. I. Theory. Foundations of Crystallography 45, 104–109 (1989)

1989
[49]

Gómez-Bombarelli, R. et al. Automatic Chemical Design Using a Data -Driven Continuous Representation of Molecules. ACS Cent. Sci. 4, 268–276 (2018)

2018
[50]

Qi, H. et al. Giant dynamic electromechanical response via field driven pseudo -ergodicity in nonergodic relaxors. Nat Commun 14, 2414 (2023)

2023
[51]

J., Tang, Y

Wang, Y. J., Tang, Y. L., Zhu, Y. L. & Ma, X. L. Entangled polarizations in ferroelectrics: A focused review of polar topologies. Acta Materialia 243, 118485 (2023)

2023
[52]

Hickox-Young, D. et al. Local structure and its implications for the relaxor ferroelectric ${\mathrm{Cd}}_{2}{\mathrm{Nb}}_{2}{\mathrm{O}}_{7}$. Phys. Rev. Res. 4, 033187 (2022)

2022
[53]

& Nelson, C

Ophus, C., Ciston, J. & Nelson, C. T. Correcting nonlinear drift distortion of scanning probe and scanning transmission electron microscopies from image pairs with orthogonal scan directions. Ultramicroscopy 162, 1–9 (2016)

2016
[54]

Malis, T., Cheng, S. C. & Egerton, R. F. EELS log -ratio technique for specimen -thickness measurement in the TEM. Journal of Electron Microscopy Technique 8, 193–200 (1988)

1988
[55]

Zhang, C., Han, R., Zhang, A. R. & Voyles, Paul. M. Denoising atomic resolution 4D scanning transmission electron microscopy data with tensor singular value decomposition. Ultramicroscopy 219, 113123 (2020)

2020
[56]

& Bokdam, M

Jinnouchi, R., Lahnsteiner, J., Karsai, F., Kresse, G. & Bokdam, M. Phase Transitions of Hybrid Perovskites Simulated by Machine -Learning Force Fields Trained on the Fly with Bayesian Inference. Phys. Rev. Lett. 122, 225701 (2019)

2019
[57]

& Kresse, G

Jinnouchi, R., Karsai, F. & Kresse, G. On -the-fly machine learning force field generation: Application to melting points. Phys. Rev. B 100, 014105 (2019)

2019
[58]

& Susi, T

Madsen, J. & Susi, T. The abTEM code: transmission electron microscopy from first principles. Open Research Europe 1, 24 (2021)

2021
[59]

& Schoelkopf, B

Tolstikhin, I., Bousquet, O., Gelly, S. & Schoelkopf, B. Wasserstein Auto -Encoders. Preprint at https://doi.org/10.48550/arXiv.1711.01558 (2019)

2019 doi arXiv
[60]

& Tanaka, I

Togo, A. & Tanaka, I. First principles phonon calculations in materials science. Scripta Materialia 108, 1–5 (2015). 17 Methods Thin film deposition. A 20 nm-thick Ba0.5Sr0.5TiO3 (BST) thin-film capacitors were grown by pulsed laser deposition (PLD) using a KrF excimer laser (...

2015

Pith tools

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

[1] [1]

Cohen, R. E. Relaxors go critical. Nature 441, 941–942 (2006)

work page 2006

[2] [2]

Yin, J. et al. Nanoscale bubble domains with polar topologies in bulk ferroelectrics. Nat Commun 12, 3632 (2021)

work page 2021

[3] [3]

Yadav, A. K. et al. Observation of polar vortices in oxide superlattices. Nature 530, 198–201 (2016)

2016

[4] [4]

Bencan, A. et al. Atomic scale symmetry and polar nanoclusters in the paraelectric phase of ferroelectric materials. Nat Commun 12, 3509 (2021)

work page 2021

[5] [5]

Pan, H. et al. Ultrahigh energy storage in superparaelectric relaxor ferroelectrics. Science 374, 100–104 (2021)

work page 2021

[6] [6]

Kumar, A. et al. Atomic-resolution electron microscopy of nanoscale local structure in lead-based relaxor ferroelectrics. Nat. Mater. 20, 62–67 (2021)

work page 2021

[7] [7]

Gong, F. -H. et al. Atomic mapping of periodic dipole waves in ferroelectric oxide. Science Advances 7, eabg5503 (2021)

work page 2021

[8] [8]

& Rappe, A

Takenaka, H., Grinberg, I., Liu, S. & Rappe, A. M. Slush -like polar structures in single -crystal relaxors. Nature 546, 391–395 (2017)

work page 2017