{"id":"a8126613-fe8d-4ac5-be10-b5d29f0d985d","arxiv_id":"2509.08321","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The cepstral method achieves 0.09% strain precision at 1 nm resolution on a small-pixel-count detector, with precision scaling proportional to convergence angle over the square root of dose.","lead":"This paper tests a signal-processing technique, the exit wave power cepstral transform, for measuring atomic-scale strain in semiconductors with low-pixel-count electron detectors. It shows the method reaches industry-relevant precision and resolution, and that precessing the electron beam improves accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Precision values and the α/√N scaling fit may be contaminated by real strain variations (thickness fringes, buckling) in the 'unstrained' Si measurement regions, as the paper's own Fig. 5 shows.","rationale":"The reader's weakest assumption precisely identifies this concern, and the paper's own Fig. 5 is direct in-text evidence that the assumption fails on wedges: the strain map shows alternating bands correlated with thickness fringes. Since Fig. 4's precision data are obtained from a Si [110] sample whose thickness was estimated to be ~110 nm—presumably on a wedge—the same artifact can inflate the local std. This is load-bearing because the paper's central quantitative deliverable is the precision number and its scaling law; if the noise floor is actually lower, the method is better than reported, but the current validation cannot distinguish. The proposed split-dataset test is feasible with existing open-source code and would settle whether the std is noise or sample-induced. This does not change the reader's CONDITIONAL verdict, but strengthens the specific condition: the authors should either measure precision on an artifact-free region or subtract common-mode strain via split-dataset analysis.","tokens_in":15318,"tokens_out":4440,"duration_ms":51773,"concrete_test":"Split the 4D-STEM dataset into two independent halves (e.g., odd/even scan lines or frames), compute strain maps from each half over the same 10x10-pixel region, and compute the standard deviation of the difference map divided by √2. Sample-related strain signals (fringes, buckling, tilt) are identical in both maps and cancel in the difference, leaving only measurement noise. If the resulting noise-only precision for the 1 pC, α=2 mrad dataset is substantially better than the reported 0.09%, the published precision values are contaminated by sample strain. Alternatively, run a multislice simulation of a perfectly unstrained Si wedge with the same thickness variation and no Poisson noise through the EWPC pipeline; if the simulated std is comparable to the reported precision, the contamination is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is the measured 0.09% precision at 1 pC (α=2 mrad) and the scaling δε ∝ α/√N (Eq. 1). Precision is quantified as the local standard deviation of strain over a ~10x10 pixel region on a nominally unstrained Si [110] sample (Fig. 4a). For this metric to measure noise, that region must contain zero real strain variation. The paper itself disproves this assumption on the same sample type: Fig. 5(b) shows strain bands in a Si wedge that exactly mirror the thickness fringes in the dark-field image, and Fig. 5(c) shows the measured 'precision' varying with thickness. If the Fig. 4 region includes even a small thickness gradient or residual buckling tilt, the std contains a systematic strain gradient that is not Poisson noise. The fitted form σ = sqrt((a/√N)^2 + b^2) cannot separate a static sample-driven strain pattern from a true noise floor; b would absorb the artifact, and the extracted a values would still be biased upward. Consequently, the headline precision numbers and the demonstrated α-scaling are not established as noise-limited, and the comparison underpinning Eq. 1 is weakened.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates the Exit Wave Power Cepstral (EWPC) transform for strain mapping with small-pixel-count 4D-STEM detectors (e.g., EMPAD). It reports precision measurements on nominally unstrained Si and on Si-SiGe multilayers, and claims a measured precision of 0.09% at 1 nm resolution for a 1 pC dose with a 2 mrad convergence semi-angle. The authors propose that precision scales as δε ∝ α/√N, with an additional constant systematic term at high dose, and demonstrate improvements from precession electron diffraction and a lack of strong improvement from zero-loss energy filtering for the studied thicknesses. The manuscript includes code availability and a systematic parameter study, but the central empirical precision values and the scaling law rest on the assumption that the chosen sample regions contain no real strain variations.","tokens_in":15667,"tokens_out":4918,"duration_ms":63050,"significance":"If the reported precision and scaling are correct, the work is significant for high-throughput semiconductor strain metrology: it directly addresses the small-pixel-count limitation of fast pixel-array detectors and offers a practical algorithm with public code. The systematic comparison across dose, convergence angle, thickness, precession, and energy filtering is valuable, and the paper honestly identifies regimes (Poisson-noise-limited vs systematic-error-limited) that are important for practical users. The strength of the paper is its empirical scope and the explicit presentation of the scaling ansatz; however, the headline precision numbers are only as strong as the assumption that the measurement regions are truly strain-free, an issue the paper itself partly exposes in Fig. 5.","major_comments":[{"comment":"The headline precision values are the local standard deviation of measured strain over a ~10x10-pixel region of a 'nominally unstrained' Si [110] foil. This metric equals measurement noise only if that region contains no real strain variation. The paper itself shows, on the same material, that thickness fringes produce alternating strain bands in EWPC strain maps (Fig. 5b) and that the local precision varies with thickness (Fig. 5c). No evidence is given that the region used for Fig. 4 is free of thickness gradients, residual tilt, buckling, or FIB damage. Any such spatially varying real strain would be absorbed into the fitted a and b parameters, biasing the claimed dose and α scalings. This is load-bearing for the central claim, so the authors should either demonstrate strain-free conditions in the exact measurement regions (e.g., with an independent measure such as a large-angle rocki","section":"Results (Fig. 4 and Fig. 5)"},{"comment":"The central scaling δε ∝ α/(θ_B√N) is presented as an analytical result from Chapman et al., but the proportionality constant is left unspecified and the experimental test of the α-dependence is a linear fit of the fitted parameter a versus α. This is a fitted trend, not a parameter-free prediction. The manuscript states the trend is 'roughly' consistent, but no uncertainties, goodness-of-fit values, or numbers of points are reported. Since this scaling is used to interpolate precision across experimental conditions and to justify the trade-off in Fig. 4(b), the fit should be reported with confidence intervals and, ideally, cross-validated against multislice simulations with known strain and dose.","section":"Eq. (1) and Supplementary Fig. 5"}],"minor_comments":[{"comment":"Eq. (1) is garbled in the text ('δε∝αθ+91N'); it should be rendered as δε ∝ α/(θ_B√N). Also, the Introduction says 'α√N scaling' where 'α/√N' is meant.","section":"Scaling of Precision with Numerical Aperture and Dose"},{"comment":"The Discussion recommends a pre-filtering step for low-dose analysis, but the headline 0.09% precision at 1 pC was obtained without one. Please clarify whether the reported number already benefits from pre-filtering or whether the recommendation post-dates the measurement.","section":"Discussion"},{"comment":"The probe-size values 'account for the finite source size', but the calculation of ds is not described. Since the resolution/precision trade-off is a central figure, a short definition or reference is needed.","section":"Fig. 4(b)"},{"comment":"The exponential fit to parameter b versus α is presented without justification. If b is intended to represent dynamical-diffraction systematic errors, a physical model or at least a discussion of alternative functional forms would strengthen the interpretation.","section":"Supplementary Text 2"},{"comment":"The size of the region around the EWPC peaks is 'manually chosen'. Since this choice can affect the fitted peak positions and hence the precision, the authors should state the actual region size used for Figs. 4-7 and comment on sensitivity.","section":"Methods (peak fitting)"},{"comment":"The paper lacks a data availability statement for the experimental datasets; only code is provided. For an empirical study of this type, making the representative raw 4D-STEM datasets available would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically sound in its algorithmic description and the parameter study is useful, but the referee report should focus on the contamination risk to the precision metric and the under-reported fit statistics. The paper would be publishable after these points are addressed, ideally with a re-analysis or a clear demonstration that the Fig. 4 regions are strain-free. I do not see grounds for rejection: the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a genuinely useful paper. The EWPC transform itself is prior work, but the systematic scaling study here is new: precision vs dose, convergence angle, and thickness on silicon, plus the first clear demonstration that precession improves cepstral strain precision by 1.3–2x and that energy filtering buys nothing for thin samples. The code is open, the experimental conditions are stated carefully, and the authors are honest about the dynamical-diffraction artifacts that persist even after the transform. I came away convinced that EWPC plus EMPAD-style detectors is a practical combination for high-throughput semiconductor strain mapping.\n\nThe main soft spot is the precision metric. They define precision as the local standard deviation over a ~10x10 pixel region of a nominally unstrained Si sample. But their own Fig. 5 shows that the same sample type can have strain bands mirroring thickness fringes, and that the measured standard deviation varies with thickness. So the quoted 0.09% at 1 pC and the fitted noise floor b are probably contaminated by real strain gradients or thickness effects. That doesn't kill the paper, but it means the precision numbers are upper bounds rather than clean noise floors, and the decomposition into Poisson vs systematic noise is not as clean as claimed. The α/√N scaling is also supported by fitted a-values, not a parameter-free prediction. The agreement with Chapman's two-pixel model is suggestive, not proof. And raw data are not deposited, only code.\n\nThese are addressable issues, not fatal ones. The right fix is to report precision on a region verified to be fringe-free (or to flatten the thickness variation), and ideally to release a small raw dataset alongside the code.\n\nWho is this for? Anyone doing nanobeam diffraction strain mapping on 4D-STEM detectors, especially in semiconductor failure analysis or process metrology. It deserves a serious referee: the experiments are careful, the claims are mostly proportionate, and the precession result is likely to be cited. I'd send it out, with a request to address the fringe contamination and to soften the claim that the precision is purely noise-limited.","headline":"Solid, useful experimental evaluation of EWPC strain mapping on small-pixel detectors; the central claim mostly holds, but the precision metric is likely contaminated by sample fringes and needs a cleaner test before the numbers are taken at face value.","tokens_in":16138,"tokens_out":1019,"would_cite":true,"duration_ms":14723,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single scaling law, δε ∝ α/√N, governs strain-map precision in few-pixel electron diffraction, and the cepstral transform reaches 0.09% precision at 1 nm resolution with a 1 pC dose.","keywords":["strain mapping","4D-STEM","cepstral transform","nanobeam electron diffraction","precession electron diffraction","pixel array detectors","Poisson noise","semiconductor metrology"],"falsifier":"Acquire repeated strain maps of the same silicon wedge after deliberate changes in specimen tilt or at locations with and without visible thickness fringes. If the local standard deviation of the strain map changes systematically with thickness or tilt, the scatter is sample-driven, not pure measurement noise. Alternatively, test the predicted √N scaling directly: at a fixed convergence angle in a verified strain-free region, quadrupling the dose should halve the Poisson-noise contribution; failure to do so would falsify the extrapolation.","tokens_in":15280,"feed_emoji":"🔬","tokens_out":5996,"duration_ms":63164,"temperature":0.7,"pith_summary":"Semiconductor metrology needs strain maps that combine 0.01–0.1% precision with nanometre resolution, but the fast pixel-array detectors best suited for high doses have small pixel counts that defeat conventional disk-tracking algorithms. This paper argues that the exit-wave power cepstral transform solves that mismatch: by Fourier-transforming the logarithm of each diffraction pattern, it converts the periodicity of many weak diffracted disks into a few sharp peaks, so precision is governed by total dose and usable collection angle rather than by the number of detector pixels. Using a two-pixel detection model, the authors derive and experimentally confirm the scaling δε ∝ α/√N, then show the trade-off between precision and spatial resolution that follows. On silicon wedges and Si–SiGe multilayers they report 0.09% strain precision at 1 nm resolution from a 1 pC, 10 ms exposure, find that precession improves precision 1.3–2.5× by suppressing dynamical diffraction artifacts, and show that energy filtering is unnecessary at device-relevant thicknesses. If correct, the result makes high-throughput, live strain mapping practical on compact detectors.","feed_headline":"Cepstral method maps strain to 0.09% precision at 1 nm","feed_subtitle":"A Fourier trick makes few-pixel diffraction patterns dose-efficient, so 10 ms exposures meet chip-industry strain targets.","key_machinery":"The load-bearing object is the Exit Wave Power Cepstral (EWPC) transform: the Fourier transform of the logarithm of the diffraction pattern, which acts as a pair-correlation function whose peaks sit at projected inter-atomic spacings. Because each cepstral peak integrates periodicity from many Bragg disks, few detector pixels and few pixels per disk suffice; what matters is the total dose N collected inside the usable angular range and the convergence semi-angle α, which enter through the two-pixel detection limit δε ∝ α/√N. The paper uses this identity to interpolate precision across experimental conditions and to locate the crossover where dynamical diffraction, rather than shot noise, set","core_discovery":"On its own terms, the claim is that strain measurement precision in nanobeam electron diffraction is fundamentally limited by Poisson counting noise and by the convergence angle, not by detector pixel count. The paper validates a cepstral analysis pipeline in which the logarithm of each convergent-beam pattern is Fourier transformed to produce pair-correlation peaks at projected inter-atomic distances; tracking two of those peaks with sub-pixel interpolation yields the full strain tensor. The measured precision follows δε ∝ α/√N in the low-dose regime, plateaus when dynamical diffraction contrast takes over at high dose, and a 0.09% precision at 1 nm resolution is achieved at 1 pC with a 2 m","pith_inferences":["If the scaling law holds at lower doses, the same pipeline could be pushed toward low-dose imaging by adding a pre-filter that suppresses diffuse-scattering noise, which the paper itself notes the logarithm can overweight.","The high-dose plateau implies that for thick or strongly scattering samples, more dose stops helping; the next lever is precession or choosing reflections less affected by dynamical contrast, not simply raising beam current.","The thickness-fringe artifacts suggest that global precision over a large field of view, not just local precision on a flat region, is the metric that will predict real-device performance; a standard strain-free wedge with documented thickness and tilt would make algorithm comparisons meaningful.","Detector developments that raise saturation current and frame rate translate directly into better precision at fixed dwell time through the √N factor, giving a quantitative target for next-generation pixel-array detectors."],"forward_implications":["A 1 pC, 10 ms-per-pixel exposure can already meet the industry's 0.1% precision / 1 nm resolution target, so production-relevant strain maps can be acquired in seconds.","Because α enters linearly, any gain in dose or in beam parallelism improves precision only at the cost of spatial resolution; the 2 mrad setting is the reported sweet spot for 0.09% at 1 nm.","Precession should be used whenever sample thickness variations or dynamical contrast dominate, improving local precision by 1.3–2× and global precision by 1.5–2.5×.","Energy filtering can be omitted for device-relevant thicknesses, removing a hardware burden unless the spectrometer itself distorts the diffraction pattern.","The method works with compact data sets and no tunable fitting parameters, making live processing and high-throughput mapping feasible."],"supporting_citations":[{"why":"Introduces the EWPC transform and its strain-mapping formalism, the core method under evaluation.","marker":"Padgett et al., 2020"},{"why":"Provides the analytic two-pixel detection limit used to derive the δε ∝ α/√N scaling law.","marker":"Chapman et al., 1978"},{"why":"Shows disk-detection precision saturates for disk diameters beyond roughly 5–6 pixels, motivating algorithms for small pixel-count detectors.","marker":"Han et al., 2018"},{"why":"Presents the high dynamic range pixel-array detector used for the fast, high-dose acquisitions.","marker":"Tate et al., 2016"},{"why":"Extends detector speed and dynamic range, enabling the short dwell times and dose-efficiency claims.","marker":"Philipp et al., 2022"},{"why":"Supplies the cross-correlation baseline software used for comparison on the same Si–SiGe dataset.","marker":"Savitzky et al., 2021"},{"why":"Introduces precession electron diffraction, the modification shown here to improve precision and suppress dynamical artifacts.","marker":"Vincent & Midgley, 1994"},{"why":"Provides the multislice simulation suite used to generate simulated diffraction patterns and thickness analysis.","marker":"Allen et al., 2015"}],"fun_headline_variants":["Cepstral strain hits 0.09% precision at 1 nm","Few-pixel detectors still achieve 0.09% strain precision","Pixel count no longer limits strain mapping precision","Cepstral method: 0.09% strain precision at 1 nm"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported precision numbers treat small regions of the nominally unstrained silicon as strain-free, so the local scatter in the strain map is counted entirely as measurement noise; real thickness fringes, buckling, or residual tilt in those regions would inflate the precision and skew the scaling fits.","fun_headline_variants_meta":{"raw":{"variants":["Cepstral strain hits 0.09% precision at 1 nm","Few-pixel detectors still achieve 0.09% strain precision","Pixel count no longer limits strain mapping precision","Cepstral method: 0.09% strain precision at 1 nm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000512,"raw_usage":{"total_tokens":2331,"prompt_tokens":752,"completion_tokens":1579,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1504}},"tokens_in":496,"tokens_out":1579,"duration_ms":10886,"temperature":1.0,"reasoning_tokens":1504,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:45:58.036282+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire repeated strain maps of the same silicon wedge after deliberate changes in specimen tilt or at locations with and without visible thickness fringes. If the local standard deviation of the strain map changes systematically with thickness or tilt, the scatter is sample-driven, not pure measurement noise. Alternatively, test the predicted √N scaling directly: at a fixed convergence angle in a verified strain-free region, quadrupling the dose should halve the Poisson-noise contribution; failure to do so would falsify the extrapolation.","supporting_citations":[],"review_version":1}