{"id":"20ae871e-8a03-4ed6-a0db-129ca14faad5","arxiv_id":"2505.24062","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Wall motion in PbTiO3 depends on the local ferroelectric-ferroelastic wall configuration, with uniform walls switching at 20 V and heterogeneous walls staying pinned until about 25 V.","lead":"This paper uses automated scanning probe microscopy to run 1,500 switching experiments on an epitaxial ferroelectric film, mapping how domain walls move under voltage pulses. It finds that the local arrangement of ferroelectric and ferroelastic domain walls controls how easily each wall moves, which matters for designing ferroelectric memory devices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The five class labels are not recoverable from the three-angle GMM assignment in Algorithm 1; if the label mapping is ambiguous or labels the wrong wall, the class-specific mobility fingerprints in Fig. 4 are not established.","rationale":"The reader's weakest assumption identifies the right area, and I sharpen it. The paper's own Algorithm 1 can only assign patches to three angle archetypes, whereas five mobility classes are used in Figures 3-4; no published rule maps one to the other. Because the GMM operates on the histogram and not on the spatially resolved orientation field, the label can come from a different wall than the one pulsed. Since the entire two-fingerprint distinction is a comparison of means across these labels, this is the single most load-bearing point. I do not see an internal mathematical contradiction beyond the under-specified label mapping, and the experimental throughput (1500 events, automated DART-PFM) is a genuine contribution. The proposed test, recomputing labels from the wall at the pulse site and from independent variant annotation, would settle whether the Figure 4 fingerprint is a property of wall classes or of the labeling heuristic. Because this is fixable with additional analysis, the conditional verdict stands without a change in disposition.","tokens_in":10860,"tokens_out":6539,"duration_ms":75982,"concrete_test":"Ask the authors to provide the explicit mapping from Algorithm 1 outputs to the five Table 1 class labels and to rerun the Figure 4 mobility analysis with labels recomputed from (i) the orientation of the ferroelastic ridge pixel nearest to the pulse location rather than the dominant GMM component of the whole 50 nm patch, and (ii) independent human or semi-automated annotation of the two domain variants on either side of the pulsed wall for a random subset of about 100 events. If the mapping is not one-to-one, if a substantial fraction of labels change, or if the Class I-III versus IV-V separation at 20 V falls below the reported error bars, the two-fingerprint claim is not robust to the classification step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests entirely on the class labels assigned to each pulse site. In the Microstructure-Resolved Switching Dynamics section and Algorithm 1, step 5, a 50 nm x 50 nm patch is summarized by a Gaussian mixture model fit to the orientation histogram of all ferroelastic ridges in the patch, and the patch label is the component with largest weight, binned to 'the nearest archetypal wall angle (0, theta, or 90 degrees)'. This cannot by itself yield the five classes of Table 1: only three angle archetypes are used, so at least two classes must share an archetype, and no rule is given for separating them. Moreover, because the orientation histogram discards spatial positions, the dominant component need not be the orientation of the ferroelastic wall actually crossed by the pulsed ferroelectric wall at the pulse site; a patch containing a kink or a nearby wall of different orientation can be labeled by a wall that was not switched. The classes are defined by pairings of domain variants (e.g., a1-/c+ vs a2-/c-), which is information not present in a single dominant orientation. Thus the Figure 4 differences between Classes I-III and IV-V could be produced by the labeling heuristic rather than by the physical wall configuration. No internal validation is reported: no manual labels, no per-wall orientation at the pulse site, and no ablated classification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":11060,"tokens_out":3864,"duration_ms":41747,"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":[{"comment":"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.","section":"Algorithm 1, step 5; Table 1"},{"comment":"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.","section":"Microstructure-Resolved Switching Dynamics; Figure 4"},{"comment":"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.","section":"Microstructure-Resolved Switching Dynamics; Figure 3"},{"comment":"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.","section":"Domain Wall Segmentation and Orientation Mapping; Algorithm 1"}],"minor_comments":[{"comment":"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.","section":"Table 1 and text, page 5"},{"comment":"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.","section":"Figure 5"},{"comment":"The phrase 'were investigated' is grammatically incorrect; 'was investigated' is needed for the singular subject 'a large area epitaxial PbTiO3 film'.","section":"Abstract, third sentence"},{"comment":"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.","section":"Data Availability"},{"comment":"The Gaussian mixture model reference is incomplete; it should include a full citation or DOI rather than a page fragment.","section":"Reference [38]"}],"recommendation":"major_revision","confidential_remarks":"The classification ambiguity is the central issue. If the authors cannot provide a clear, validated class-assignment rule, the paper's main physical claim—the distinct mobility fingerprints—is not established. The internal inconsistency between the text and Table 1 suggests the manuscript may have been prepared hastily. I would urge the editor to require the missing labeling rule, statistical tests, and sensitivity analysis before considering publication. The topic fits the journal's scope, and the automation contribution is real, but the evidence needs to match the strength of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.","headline":"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.","tokens_in":11655,"tokens_out":1151,"would_cite":false,"duration_ms":14114,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["77.80.Fm","68.37.Ps"],"model":"deepseek-v4-flash","headline":"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…","keywords":["domain wall pinning","ferroelectric switching","ferroelastic domains","piezoresponse force microscopy","high-throughput AFM","machine learning classification","PbTiO3 thin films","domain wall mobility"],"falsifier":"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.","tokens_in":10592,"feed_emoji":"🔬","tokens_out":8290,"duration_ms":75473,"temperature":0.7,"pith_summary":"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.","feed_headline":"Two wall-switching fingerprints found in PbTiO3 by automated AFM","feed_subtitle":"Uniform ferroelastic walls switch near 20 V; mixed walls pin until ~25 V, then jump.","key_machinery":"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 $30\\,\\mathrm{V}$.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the dual-frequency resonance-tracking PFM mode used to acquire the amplitude and phase maps.","marker":"34"},{"why":"Provides the Sato ridge filter that enhances elongated ferroelastic wall features in the amplitude channel.","marker":"36"},{"why":"Provides the Canny edge detection used to isolate 180-degree ferroelectric boundaries from the phase channel.","marker":"37"},{"why":"Provides the Gaussian mixture model used to classify each patch's dominant ferroelastic orientation.","marker":"38"},{"why":"Establishes the Bayesian-optimization automated SPM framework that motivates the machine-learning-controlled high-throughput workflow.","marker":"25"},{"why":"Demonstrates automated bias-pulse creation of metastable domain states in PFM, which the pulse protocol here builds on.","marker":"27"},{"why":"Supplies the closed-loop hypothesis-learning approach that the paper positions its high-throughput experiment against.","marker":"31"}],"fun_headline_variants":["Automated AFM reveals two wall-switching regimes in PbTiO3","Wall geometry dictates switch bias in PbTiO3 ferroelectric","Pinning classes split by ferroelastic orientation in PbTiO3","Two switching fingerprints for ferroelectric walls found via AFM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Automated AFM reveals two wall-switching regimes in PbTiO3","Wall geometry dictates switch bias in PbTiO3 ferroelectric","Pinning classes split by ferroelastic orientation in PbTiO3","Two switching fingerprints for ferroelectric walls found via AFM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000596,"raw_usage":{"total_tokens":2851,"prompt_tokens":1068,"completion_tokens":1783,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":1710}},"tokens_in":684,"tokens_out":1783,"duration_ms":11827,"temperature":1.0,"reasoning_tokens":1710,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:36:23.966361+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}