Exploring Domain Wall Pinning in Ferroelectrics via Automated High Throughput AFM

Hiroshi Funakubo; Kamyar Barakati; Sergei V. Kalinin; Yu Liu

REVIEW 4 major objections 5 minor 1 cited by

Exploring Domain Wall Pinning in Ferroelectrics via Automated High Throughput AFM

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 In epitaxial PbTiO3, a ferroelectric domain wall's response to a voltage pulse is set by its local ferroelastic wall class: uniform walls switch coherently above about 20 V, while heterogeneous walls remain pinned until roughly 25 V and…

desk verdict A useful automated PFM workflow and a new 1500-event dataset, but the central claim about class-specific mobility fingerprints rests on a labeling scheme that the paper never actually defines. read the letter →

arxiv 2505.24062 v1 pith:3CGNEBNW submitted 2025-05-29 cond-mat.mtrl-sci cs.CVcs.LGphysics.app-ph

Kamyar Barakati , Yu Liu , Hiroshi Funakubo , Sergei V. Kalinin This is my paper · ORCID

classification cond-mat.mtrl-scics.CVcs.LGphysics.app-ph PACS 77.80.Fm68.37.Ps

keywords domainwallpinningferroelectricswitchingferroelasticdomainspiezoresponseforcemicroscopyhigh-throughputAFMmachinelearningclassificationPbTiO3thinfilmsmobility

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 reports that, in an epitaxial $\mathrm{PbTiO}_3$ film on $\mathrm{KTaO}_3$, whether a pre-existing ferroelectric domain wall moves under a voltage pulse is governed by the local ferroelastic wall class around it. Using a machine-learning-controlled piezoresponse force microscope that executed 1500 switching events in about 30 hours, the authors find two distinct mobility fingerprints: uniform ferroelastic walls (three of the five classes) stay nearly static at $10\,\mathrm{V}$ and then jump coherently once the bias reaches about $20\,\mathrm{V}$, while heterogeneous walls remain pinned at $20\,\mathrm{V}$ and only release near $25\,\mathrm{V}$, after which they displace rapidly. The contribution is a scalable measurement strategy that turns sparse, human-selected PFM experiments into a statistically dense map of structure-property rules, connecting wall orientation to the critical pulse voltage. The authors argue these rules can serve as a design basis for ferroelectric memory devices.

What carries the argument

The carrying object is the patch-level wall class, computed by a five-stage image pipeline. A DART-PFM amplitude map is enhanced with a Sato ridge filter; a structure-tensor computation converts the ridges into a continuous orientation field; Canny edge detection on the phase channel isolates 180-degree ferroelectric walls, which are then masked out to leave only ferroelastic orientations. Around each of the 1500 pulse sites, a $50 \times 50\,\mathrm{nm}^2$ patch is histogrammed in orientation, and a Gaussian mixture model fit to that histogram yields a dominant orientation that is snapped to one of five archetypal wall classes (I-V). The measurement that carries the argument is the pixel-wise absolute difference between pre- and post-pulse amplitude and phase images, compared across the five classes at $10\,\mathrm{V}$, $20\,\mathrm{V}$, and $30\,\mathrm{V}$.

What would settle it

Classify the same pulse sites two independent ways, via the paper's orientation histogram and Gaussian mixture model and via high-resolution reciprocal-space mapping or cross-sectional electron microscopy of the identical walls, and check that the $20\,\mathrm{V}$ versus $25\,\mathrm{V}$ activation difference survives the relabeling; or run pulses at intermediate voltages such as $22.5\,\mathrm{V}$ and $27.5\,\mathrm{V}$ to see whether Class IV-V displacement really plateaus below $25\,\mathrm{V}$ and jumps afterward, as the two-step fingerprint predicts.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the ferroelectric-ferroelastic geometry of a wall segment determines how it responds to an applied bias. Class I-III walls, where the 180-degree ferroelectric trace cuts between ferroelastic domains aligned along the same crystallographic direction, show a single coherent activation: displacement stays below about 15 percent of the patch area at $10\,\mathrm{V}$, rises to roughly 20-25 percent at $20\,\mathrm{V}$, and reaches 26-28 percent at $30\,\mathrm{V}$. Class IV-V walls, where the ferroelectric wall separates ferroelastic domains of different orientations, are pinned at both $10\,\mathrm{V}$ and $20\,\mathrm{V}$ and only above about $25\,\mathrm{V}$ give way, jumping to more than 20 percent displacement at $30\,\mathrm{V}$. The authors further report that at $30\,\mathrm{V}$ wall mobility is angularly structured, with maxima near $50^\circ$ and $140^\circ$ and minima near $40^\circ$, $80^\circ$, and $155^\circ$, which they interpret as signatures of polarization frustration or strain incompatibility at particular wall geometries.

Load-bearing premise

The load-bearing premise is that a $50 \times 50\,\mathrm{nm}^2$ patch, labeled by the dominant ferroelastic orientation in its histogram, truly represents the local wall configuration that controls switching, and that the pixel-wise pre/post image difference is an uncontaminated measure of wall displacement.

Editorial extensions

If this is right

For memory design, the class labels give a direct rule: write pulses near $20\,\mathrm{V}$ will move uniform walls but not heterogeneous ones, so pulse margins can be tuned to avoid crosstalk between wall types.
The 10-fold throughput gain (1500 events in 30 hours versus roughly 150 by hand) makes wall-class-resolved statistics practical in other epitaxial films.
The angular mobility map at $30\,\mathrm{V}$ implies that some wall orientations stay pinned even at the highest field used here, allowing orientation maps to act as predictive switching maps.
The same automated pipeline, extended with EBSD data, could transfer the class taxonomy to polycrystalline films and grain boundaries.

Reading between the lines

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

A direct time-resolved extension, imaging the same wall during the pulse rather than before and after, would test whether the creep-and-unpin language describes genuine creep below $25\,\mathrm{V}$ or simply a threshold jump; the present data cannot distinguish them.
The five-class scheme is fit to one composition and strain state; applying it to other epitaxial systems would show whether the $20\,\mathrm{V}$/$25\,\mathrm{V}$ thresholds are universal to the geometry or specific to $\mathrm{PbTiO}_3/\mathrm{KTaO}_3$.
The angular maxima and minima suggest a practical design knob: choosing the pulse location's wall-trace orientation relative to the crystallographic axes should either maximize or suppress wall motion at a given voltage, which could be tested by targeted pulses on selected orientations.

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 actual new thing here is the pipeline and the dataset: 1500 paired pre/post DART-PFM switching events on epitaxial PTO/KTO, collected automatically in ~30 hours with clear descriptions of the Sato filtering, Canny masking, and GMM-based patch classification. That is a real contribution. The empirical observation that switching depends on local wall geometry is plausible, and the qualitative distinction between uniform and heterogeneous ferroelastic walls is a reasonable extension of known pinning physics. Credit where due: the workflow is detailed enough to reproduce, and the 1500-event statistics are better than the handful of manual experiments the field usually sees. The central claim, however, is not established. The five classes of Table 1 are defined by pairings of domain variants (a1-/c+ vs a2-/c- etc.), but Algorithm 1 assigns each patch a single dominant angle from a GMM fit to the ridge orientation histogram, then bins that angle to one of three archetypes: 0°, θ°, or 90°. Nowhere is it explained how those three angles map onto the five classes, and because the orientation histogram discards spatial position, the dominant component does not have to be the wall that was actually pulsed. The Figure 4 fingerprints could therefore be produced by the labeling heuristic rather than by the physical wall configuration. This is the load-bearing soft spot. There are also smaller issues: no significance tests separate the classes (the error bars overlap in ways that look non-significant between some groups), the 'above ~25 V' claim is interpolated from only 10, 20, and 30 V data, and the data/code are not public despite the paper leaning on the 'high-throughput' selling point. The class descriptions in the text and Table 1 are inconsistent (Class I vs Class II both list a2+/c- || a2-/c+ with only an angle difference). None of these flaws make the workflow useless; they just mean the headline result needs more work. A serious referee should ask for a clearly defined label mapping, per-pulse orientation validation or manual labels, significance tests, and either actual 25 V data or softened wording. If the authors provide that, this could be a solid methods paper. As is, I would not yet build anything on the class-specific fingerprints. Send it to review, but with the expectation of substantial revision.

Referee Report

4 major / 5 minor

Summary. The manuscript reports an automated Piezoresponse Force Microscopy (PFM) workflow in which a machine-learning-controlled controller selects 1500 pulse sites on epitaxial PbTiO3/KTaO3, applies 10-30 V pulses, and measures domain-wall displacement from pre/post-pulse image differences. Walls are grouped into five classes based on local ferroelastic orientation in 50 nm patches, and the paper claims two mobility fingerprints: uniform ferroelastic walls (Classes I-III) switch coherently above a single threshold near 20 V, while heterogeneous walls (Classes IV-V) remain pinned until roughly 25 V and then displace rapidly. The paper also presents an angular dependence of wall mobility at 30 V and a roadmap for autonomous multimodal SPM.

Significance. If the central claim holds, the work would be a valuable demonstration of high-throughput automated AFM, with a dataset of 1500 switching events and a seemingly generalizable pipeline for microstructure-resolved switching studies. The authors provide substantial detail on the imaging, pulse, and segmentation pipeline, and the claimed 10-fold throughput increase over manual experimentation is notable. However, the scientific value is entirely contingent on the validity of the five-class labeling and the displacement metric; as written, the classification rule is under-specified and the statistical basis for the fingerprints is missing. The paper's strengths are its scale and automation, but the load-bearing analysis needs to be established with a clear, validated classification and with significance tests.

major comments (4)

[Algorithm 1, step 5; Table 1] The manuscript does not define how the five classes are recovered from the patch-orientation pipeline. Algorithm 1 assigns each patch to the nearest archetypal angle (0°, θ°, or 90°), but the table and text list five classes, with Classes IV and V not associated with any distinct angle. No rule is given to separate classes that share the same archetype, and the text and Table 1 disagree on whether Class II is orthogonal at 90° or inclined at θ°. Consequently, the class labels used in Figure 4 are not reproducible from the described algorithm. Because the central claim compares Classes I-III against IV-V, the authors must provide an unambiguous labeling rule and validate it, for example with manual labels or the orientation of the specific ferroelastic wall at the pulse site.
[Microstructure-Resolved Switching Dynamics; Figure 4] The claimed two mobility fingerprints are not supported by statistical inference. Figure 4 shows mean pixel displacement with one-standard-deviation error bars, but no significance tests are reported for the class-by-voltage differences. The statement that Classes IV-V 'remain largely static at both 10 V and 20 V' and then undergo 'rapid displacement to >20% at 30 V' requires at least pairwise comparisons or bootstrap confidence intervals. Moreover, the '~25 V' threshold is an interpolation between the 20 V and 30 V measurements; no measurement or model at 25 V is presented, so the two-step creep-and-unpin description is an interpretive curve rather than an observed threshold.
[Microstructure-Resolved Switching Dynamics; Figure 3] The pixel-wise absolute difference between pre- and post-pulse amplitude and phase images is asserted to be a rigorous measure of domain-wall displacement, but no control experiments are reported that would separate true wall motion from scanning drift, tip wear, topographic changes, or non-switching contrast variations. Without a noise-floor estimate (e.g., repeated scans without a voltage pulse), the magnitude thresholds used to classify 'static' versus 'activated' walls are not validated. Please provide null-experiment statistics or an independent metric (e.g., direct wall-position tracking) to calibrate the difference field.
[Domain Wall Segmentation and Orientation Mapping; Algorithm 1] The classification depends on several tunable parameters—patch side length s, Sato filter parameters α and β and the scale range, the 90th-percentile threshold, Canny edge thresholds, and the GMM component-selection criterion. The paper reports no sensitivity analysis for any of these choices. Since the class labels are the independent variable in the central comparison, the authors should show that the Figure 4 fingerprints are stable under reasonable variations of these parameters, or at least quantify how label noise propagates to the displacement statistics.

minor comments (5)

[Table 1 and text, page 5] There is an internal inconsistency: the text states Class II walls are 'orthogonal... perpendicular to principal planes' and Class III walls are 'angled at θ', while Table 1 assigns Class II to 'angled at θ°' and Class III to 'angled at 90°'. These definitions should be reconciled.
[Figure 5] The angular trend at 30 V, with maxima near 50° and 140° and minima near 40°, 80°, and 155°, is presented without error bars or statistical tests; either add confidence intervals or soften the claim to a qualitative observation.
[Abstract, third sentence] The phrase 'were investigated' is grammatically incorrect; 'was investigated' is needed for the singular subject 'a large area epitaxial PbTiO3 film'.
[Data Availability] The data availability statement says data are available 'upon request' but does not mention code or analysis scripts; given the emphasis on automated and ML-controlled workflows, releasing the analysis pipeline would aid reproducibility.
[Reference [38]] The Gaussian mixture model reference is incomplete; it should include a full citation or DOI rather than a page fragment.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the mobility fingerprints are measured independently of the orientation labels; self-citations are descriptive, not load-bearing.

full rationale

The derivation chain is not circular. The five wall-orientation classes are defined from pre-pulse PFM orientation histograms via Sato filtering, structure-tensor orientation mapping, Canny phase-edge masking, and GMM fitting (Algorithm 1). The dependent variable is the pixel-wise absolute difference between pre- and post-pulse PFM images, defined in the Microstructure-Resolved Switching Dynamics section, and is computed without using the GMM labels or any fitted parameter. The conclusion that Classes I-III exhibit a single threshold while Classes IV-V show creep-and-unpin is therefore an empirical association between an independently measured response and the orientation-based labels, not a quantity forced by construction. The nearest-archetype assignment (0 degrees, theta degrees, 90 degrees) is under-specified relative to the five classes in Table 1, and the absence of a held-out prediction means the 'predictive map' is prospective rather than demonstrated; these are validity concerns, not circularity. The numerous self-citations (refs 25-31, 34) describe prior automated PFM and DART-PFM tools used as background or method components; they do not supply the paper's central claim, which rests on the 1500 paired measurements and the image-processing pipeline described in the text.

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

No new physical entities are introduced. The central claim rests on the domain-wall classification from image processing and on the assumption that pixel difference measures wall displacement. Analysis hyperparameters, such as patch size and threshold percentiles, are chosen without sensitivity checks.

free parameters (4)

Patch side length s = 50 nm
Sets the neighborhood over which each pulse site is assigned a ferroelastic orientation; changing it changes the class label and therefore the mobility statistics.
Sato filter parameters alpha, beta, scale range = not stated
Controls ridge enhancement and orientation map quality in the amplitude channel; chosen by hand without a reported sensitivity study.
Orientation-response percentile threshold = 90th percentile
Isolates reliable orientation vectors; the threshold choice affects which pixels contribute to the orientation histograms.
Canny edge thresholds = not stated
Sets the ferroelectric wall mask; affects which pixels are excluded from the ferroelastic orientation map.

assumptions (4)

domain assumption The mixed a1/a2 twin and c-lamella domain hierarchy in epitaxial PbTiO3/KTaO3 forms as described and is stable during imaging.
The whole interpretation relies on the material-level description from refs 11-13, not verified within this paper.
domain assumption PFM phase channels unambiguously mark 180 degree ferroelectric walls, and amplitude ridges mark ferroelastic walls.
Used in Algorithm 1 and Figure 2; if phase and amplitude contrast mix wall types, the separation of ferroelectric and ferroelastic walls fails.
domain assumption A 50 nm patch centered on a wall is large enough to characterize the local ferroelastic configuration and small enough to isolate a single wall.
Patch size is chosen without a sensitivity analysis, despite being central to class assignment.
standard math The Gaussian mixture model with component count chosen by BIC captures the true orientation distribution in each patch.
Model selection by BIC is standard, but the optimum depends on binning and patch content.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Domain Wall Pinning in Ferroelectrics via Automated High Throughput AFM." pith.science (2026). https://pith.science/paper/3CGNEBNW

@misc{pith2026250524062,
  author       = {Pith},
  title        = {Pith review of: Exploring Domain Wall Pinning in Ferroelectrics via Automated High Throughput AFM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CGNEBNW}},
  note         = {Machine review of arXiv:2505.24062}
}

abstract

Domain-wall dynamics in ferroelectric materials are strongly position-dependent since each polar interface is locked into a unique local microstructure. This necessitates spatially resolved studies of the wall-pinning using scanning-probe microscopy techniques. The pinning centers and preexisting domain walls are usually sparse within image plane, precluding the use of dense hyperspectral imaging modes and requiring time-consuming human experimentation. Here, a large area epitaxial PbTiO$_3$ film on cubic KTaO$_3$ were investigated to quantify the electric field driven dynamics of the polar-strain domain structures using ML-controlled automated Piezoresponse Force Microscopy. Analysis of 1500 switching events reveals that domain wall displacement depends not only on field parameters but also on the local ferroelectric-ferroelastic configuration. For example, twin boundaries in polydomains regions like a$_1^-$/$c^+$ $\parallel$ a$_2^-$/$c^-$ stay pinned up to a certain level of bias magnitude and change only marginally as the bias increases from 20V to 30V, whereas single variant boundaries like a$_2^+$/$c^+$ $\parallel$ a$_2^-$/$c^-$ stack are already activated at 20V. These statistics on the possible ferroelectric and ferroelastic wall orientations, together with the automated, high-throughput AFM workflow, can be distilled into a predictive map that links domain configurations to pulse parameters. This microstructure-specific rule set forms the foundation for designing ferroelectric memories.

Figures

Figures reproduced from arXiv: 2505.24062 by the authors.

**Figure 4.** Dependence of ferroelastic wall displacement on pulse magnitude for each orientation cluster. Mean pixel‐wise absolute displacement (expressed as a percentage of the 50 nm × 50 nm patch area) for each of the five wall‐orientation clusters after 10 𝑠 pulses at 10 𝑉, 20 𝑉, and 30 𝑉. Error bars denote one standard deviation across all sampled patches. At high field 30 𝑉, the angular trend shown in [PITH_FULL_IMAGE:fig… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Domain Switching on the Pareto Front: Multi-Objective Deep Kernel Learning in Automated Piezoresponse Force Microscopy
cond-mat.mtrl-sci 2025-06 conditional novelty 6.0 of 10

A multi-objective active-learning microscope workflow maps ferroelectric domain-switching Pareto frontiers and predicts switching ease from structural images alone.

Reference graph

Works this paper leans on

5 extracted references · 4 canonical work pages · cited by 1 Pith paper

[3]

ferroelectricity

Physical Review B 2002, 65 (10), 104111. (19) Seidel, J.; Martin, L. W.; He, Q.; Zhan, Q.; Chu, Y .-H.; Rother, A.; Hawkridge, M.; Maksymovych, P.; Yu, P.; Gajek, M. Conduction at domain walls in oxide multiferroics. Nature materials 2009, 8 (3), 229-234. (20) Paruch, P.; Giamarchi, T.; Tybell, T.; Triscone, J.-M. Nanoscale studies of domain wall motion i...

arXiv 2002
[4]

Dependence of ferroelastic wall displacement on pulse magnitude for each orientation cluster. Mean pixel‐wise absolute displacement (expressed as a percentage of the 50 nm × 50 nm patch area) for each of the five wall‐orientation clusters after 10 𝑠 pulses at 10 𝑉, 20 𝑉, and 30 𝑉. Error bars denote one standard deviation across all sampled patches. At hig...

work page 2006
[5]

R.; Rappe, A

(39) Xu, R.; Liu, S.; Grinberg, I.; Karthik, J.; Damodaran, A. R.; Rappe, A. M.; Martin, L. W. Ferroelectric polarization reversal via successive ferroelastic transitions. Nature materials 2015, 14 (1), 79-86. (40) RESEARCH, A. Ue Vero Interferometric AFM. https://afm.oxinst.com/assets/uploads/Asylum%20Research/App%20Notes/oe%20Vero%20Interferometric%20AF...

work page 2015
[8]

Applied Physics Letters 2012, 100 (23)

O3 barriers. Applied Physics Letters 2012, 100 (23). (7) Choi, K. J.; Biegalski, M.; Li, Y .; Sharan, A.; Schubert, J.; Uecker, R.; Reiche, P.; Chen, Y .; Pan, X.; Gopalan, V . Enhancement of ferroelectricity in strained BaTiO3 thin ﬁlms. Science 2004, 306 (5698), 1005-1009. (8) Liu, Y .; Kelley, K. P.; Vasudevan, R. K.; Zhu, W.; Hayden, J.; Maria, J. P.;...

work page 2012
[2024]

J.; Callahan, C.; Kalinin, S

(34) Rodriguez, B. J.; Callahan, C.; Kalinin, S. V .; Proksch, R. Dual-frequency resonance-tracking atomic force microscopy. Nanotechnology 2007, 18 (47). DOI: Artn 475504 10.1088/0957-4484/18/47/475504. (35) Jesse, S.; Guo, S.; Kumar, A.; Rodriguez, B.; Proksch, R.; Kalinin, S. V . Resolution theory, and static and frequency-dependent cross-talk in piezo...

work page doi:10.1088/0957-4484/18/47/475504 2007

Pith tools

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

[1] [3]

ferroelectricity

Physical Review B 2002, 65 (10), 104111. (19) Seidel, J.; Martin, L. W.; He, Q.; Zhan, Q.; Chu, Y .-H.; Rother, A.; Hawkridge, M.; Maksymovych, P.; Yu, P.; Gajek, M. Conduction at domain walls in oxide multiferroics. Nature materials 2009, 8 (3), 229-234. (20) Paruch, P.; Giamarchi, T.; Tybell, T.; Triscone, J.-M. Nanoscale studies of domain wall motion i...

arXiv 2002

[2] [4]

Dependence of ferroelastic wall displacement on pulse magnitude for each orientation cluster. Mean pixel‐wise absolute displacement (expressed as a percentage of the 50 nm × 50 nm patch area) for each of the five wall‐orientation clusters after 10 𝑠 pulses at 10 𝑉, 20 𝑉, and 30 𝑉. Error bars denote one standard deviation across all sampled patches. At hig...

work page 2006

[3] [5]

R.; Rappe, A

(39) Xu, R.; Liu, S.; Grinberg, I.; Karthik, J.; Damodaran, A. R.; Rappe, A. M.; Martin, L. W. Ferroelectric polarization reversal via successive ferroelastic transitions. Nature materials 2015, 14 (1), 79-86. (40) RESEARCH, A. Ue Vero Interferometric AFM. https://afm.oxinst.com/assets/uploads/Asylum%20Research/App%20Notes/oe%20Vero%20Interferometric%20AF...

work page 2015

[4] [8]

Applied Physics Letters 2012, 100 (23)

O3 barriers. Applied Physics Letters 2012, 100 (23). (7) Choi, K. J.; Biegalski, M.; Li, Y .; Sharan, A.; Schubert, J.; Uecker, R.; Reiche, P.; Chen, Y .; Pan, X.; Gopalan, V . Enhancement of ferroelectricity in strained BaTiO3 thin ﬁlms. Science 2004, 306 (5698), 1005-1009. (8) Liu, Y .; Kelley, K. P.; Vasudevan, R. K.; Zhu, W.; Hayden, J.; Maria, J. P.;...

work page 2012

[5] [2024]

J.; Callahan, C.; Kalinin, S

(34) Rodriguez, B. J.; Callahan, C.; Kalinin, S. V .; Proksch, R. Dual-frequency resonance-tracking atomic force microscopy. Nanotechnology 2007, 18 (47). DOI: Artn 475504 10.1088/0957-4484/18/47/475504. (35) Jesse, S.; Guo, S.; Kumar, A.; Rodriguez, B.; Proksch, R.; Kalinin, S. V . Resolution theory, and static and frequency-dependent cross-talk in piezo...

work page doi:10.1088/0957-4484/18/47/475504 2007