{"id":"1cd2ded9-f5fb-40c2-b180-4b0150e3c65a","arxiv_id":"2501.09743","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A preprocessing pipeline (drift alignment plus intensity transforms) enables PCA and k-means clustering to compare RHEED videos from different LaFeO3 growths.","lead":"This paper adds image-alignment and contrast-enhancement steps to electron diffraction videos from thin-film growth, which lets machine learning compare different growth runs. The authors demonstrate it on two perovskite oxide films, a step toward automating the search for good growth conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RSS alignment's bilateral-symmetry assumption is untested across azimuths and growth stages; if violated, cross-sample PCA/k-means may reflect alignment bias rather than physical film differences.","rationale":"The paper's central claim is that RSS alignment mitigates pattern translation so PCA/k-means focus on physical features, enabling the first multi-sample comparison. The load-bearing assumption is the geometric symmetry used to define the alignment. The reader's weakest_assumption identified exactly this. I agree: the synthetic drift test is strong evidence for the specific pattern used, but it does not test symmetry robustness. The azimuth test I propose would directly falsify or validate the assumption. If asymmetry causes bias, the cross-sample comparisons in Figures 6–7 could be artifacts. The novelty claim about 'first demonstration' is also questionable given ref 14, but that affects priority, not correctness. The symmetry assumption affects correctness, so it is more load-bearing. The reader's verdict of CONDITIONAL already captures this; my analysis does not move it. Credit is due for the open data, the clear synthetic-drift demonstration, and the honest acknowledgement that alignment works best on clean substrates with sharp features.","tokens_in":8582,"tokens_out":6397,"duration_ms":67678,"concrete_test":"Record RHEED videos from a static, well-oriented LaFeO3/Nb:SrTiO3 sample (or a clean Nb:SrTiO3 substrate) with no stage motion at several azimuthal angles, e.g., 0°, 5°, and 10° off the zone axis, keeping the electron beam fixed. Apply the RSS alignment algorithm to each video and extract the recovered horizontal and vertical centers for each frame. If the recovered center shifts by more than 2 pixels between azimuths (or drifts within a video) while the physical specular spot position is constant, the bilaterality assumption is violated and cross-sample comparisons are unreliable. Complement with a synthetic test: take a real RHEED image, add a known translation plus an asymmetric Kikuchi-like feature, and check whether RSS alignment's recovered translation is biased by more than 1 pixel.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II.B's RSS alignment assumes bilateral symmetry of RHEED patterns across the vertical axis and horizontal symmetry of the specular spot. The entire cross-sample PCA/k-means pipeline (Section III.B, Figures 6–7) depends on this assumption: any asymmetry – from off-azimuth alignment, vicinal surfaces, Kikuchi band contrast, shadow edges, or refraction effects – biases the recovered center. Because each frame is aligned independently, a pattern-dependent asymmetry during growth or between samples would be misinterpreted as drift or physical difference, directly corrupting the eigenvalues, eigenvectors, and clusters that the central claim says 'inform on the film surface quality.' The only validation (Figure 4) is a single video with synthetic constant-rate translation; it does not probe symmetry variation across azimuths, terminations, or growth stages. The 'within a couple of pixels' cross-sample accuracy is asserted without an error analysis or ground-truth center comparison, leaving the load-bearing assumption unquantified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes and demonstrates a preprocessing pipeline for RHEED video data to enable unsupervised PCA and k-means clustering across different MBE-grown LaFeO3 samples. The pipeline comprises frame cropping, intensity transformations (piecewise, power, inverse power), and a residual-sum-of-squares (RSS) alignment procedure that uses mirror symmetry to identify the specular spot and correct for pattern translation. The authors show on a single LFO growth that a synthetic constant-rate drift changes PCA/k-means results and that RSS alignment nearly restores the original results. They then apply the pipeline to two LFO films grown at nominally identical conditions, reporting common initial clusters and subsequent divergence, and interpret this as the first cross-sample quantitative RHEED comparison.","tokens_in":8723,"tokens_out":5259,"duration_ms":54066,"significance":"If the claims are fully supported, the work would be a useful methods contribution to RHEED-based machine learning: it addresses a real preprocessing problem, uses unsupervised methods without fitting to a desired output, ships openly available data, and provides a controlled synthetic-drift test that cleanly demonstrates the effect of translation on PCA/k-means. The power-transform robustness check is also a positive feature, since it shows that the unsupervised results are not forced by the intensity rescaling choice. However, the central cross-sample claim rests on the symmetry assumption of the alignment algorithm and on mostly qualitative validation, so the significance is contingent on additional quantitative tests.","major_comments":[{"comment":"The RSS alignment algorithm assumes bilateral symmetry of RHEED patterns about the vertical axis and horizontal symmetry of the specular spot. This assumption is load-bearing for the cross-sample comparison in Figures 6 and 7, but it is tested only against a synthetic constant-rate translation of a single symmetric pattern (Figures 4 and 5). If the symmetry is broken by off-azimuth alignment, vicinal surface steps, Kikuchi-band contrast asymmetries, shadow edges, or refraction, the mirror-based RSS minimization will bias the recovered center, and because each frame is aligned independently, a pattern-dependent asymmetry during growth would be encoded as drift or as a physical difference. The authors should either validate the symmetry assumption across azimuths and growth stages or quantify the alignment bias and show that it is small compared with the physical differences they report.","section":"II.B, III.B"},{"comment":"The assertion that RSS alignment aligns recordings from different samples 'with accuracy within a couple of pixels' is not backed by any error analysis or ground-truth comparison. No metric is reported for the residual alignment error on real data, and no independent check (e.g., tracking a stationary feature, comparing with manual alignment, or measuring the post-alignment scatter of the specular spot position) is provided. Without such a metric, the 'quantitative comparison' claim in Section III.B and the Conclusion is not established, because residual misalignment could contribute to the eigenvalue and cluster differences observed in Figures 6 and 7.","section":"III.B"},{"comment":"The cross-sample comparison is interpreted qualitatively: cluster labels are assigned by visual inspection of centroid images, and the divergence between growths is described narratively. The authors should add a quantitative measure of sample-to-sample similarity in PCA space or cluster occupancy (e.g., distances between sample trajectories in the first few principal components, cluster centroid correlations, or a confusion-style comparison of cluster assignments) to support the claim that the pipeline 'allows for quantitative comparison of RHEED videos' rather than merely a side-by-side visual display.","section":"III.B, Figures 6-7"}],"minor_comments":[{"comment":"There is a typo in 'osciallations'; the sentence would also read more smoothly as 'recognizing that the sample has become amorphous or that large islands have formed on the surface.'","section":"Abstract"},{"comment":"'outline in 12' should be 'outlined in Ref. 12', and 'h5Ppy' should likely be 'h5py'.","section":"II.C"},{"comment":"'Due to the power recalling function' should read 'Due to the power rescaling function'.","section":"III.A"},{"comment":"The text says panels (a-c) and (g-h) are raw and (d-f) and (j-l) are transformed, but the caption lists (a-f) as eigenvalues and (g-l) as eigenvectors; this leaves panel (i) unaccounted for and should be corrected.","section":"III.A, Figure 2"},{"comment":"'procedes' should be 'proceeds', and 'the development of the films surface' should be 'the development of the film's surface'.","section":"III.B"},{"comment":"The low-intensity filter threshold at 90% of maximum spot intensity is introduced without a sensitivity analysis; a sentence justifying this value or showing robustness would help reproducibility.","section":"II.B"},{"comment":"The data repository is cited, but no mention is made of code availability; sharing the preprocessing code would aid reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is a methods contribution with a modest but useful scope. My main reservation is that the load-bearing symmetry assumption and the cross-sample accuracy claim are insufficiently validated. A focused error-analysis test on the existing data, such as measuring the post-alignment scatter of the specular spot position or perturbing the initial crop and checking stability, should be feasible and would substantially strengthen the manuscript. The 'first demonstration' claim may also need to be weighed against Ref. 14, which already performs comparative dimension reduction across samples."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The RSS alignment algorithm is the real contribution here, and the synthetic drift test is the strongest part: after alignment, PCA and k-means on a drifted video match the original, which is exactly the evidence you need for a registration method. Open data is a plus, and the paper is honest that the intensity transforms mostly change contrast without altering the clustering.\n\nThe soft spots are real but not fatal. The bilateral-symmetry assumption is load-bearing, and the paper only tests it on one video with artificial translation. That test cannot detect a systematic center bias caused by asymmetric patterns, which could plausibly happen with off-azimuth alignment, shadow edges, or Kikuchi band asymmetry. The stress-test concern about symmetry lands. The cross-sample demonstration is only two growths, and 'within a couple of pixels' is given without an error analysis. I would want ground truth or at least a sensitivity check on the symmetry assumption before treating the cross-sample PCA/k-means as quantitative. The 'first demonstration' claim also needs to be phrased carefully relative to Gliebe and Sehirlioglu (ref 14); if that work already did comparative dimension reduction, the novelty has to be the RSS alignment specifically, not the cross-sample idea.\n\nNone of this undermines the core method. The paper deserves a serious referee. The preprocessing details and open data will be useful to anyone building RHEED-based ML for MBE. The right outcome is publication after revisions: validate the symmetry assumption or state its limits, replace 'within a couple pixels' with a real error estimate, and tighten the novelty language. I would cite it if I worked on RHEED analytics.","headline":"A useful, mostly sound RHEED preprocessing paper: the RSS drift-correction is new and cleanly demonstrated, but the cross-sample claims rest on an untested symmetry assumption and a qualitative two-growth comparison.","tokens_in":9315,"tokens_out":1875,"would_cite":true,"duration_ms":21605,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mirror-based drift correction makes RHEED video analysis quantitative, enabling cross-sample machine-learning comparison for the first time.","keywords":["RHEED","principal component analysis","k-means clustering","drift correction","molecular beam epitaxy","LaFeO3","machine learning","in situ characterization"],"falsifier":"Record a RHEED video at an azimuth where the pattern is known to be asymmetric, apply RSS alignment, and compare the recovered center against the known true pattern location: if the recovered center shifts by more than a pixel when no motion is applied, or fails to track an imposed pixel offset, the symmetry premise fails. A simpler synthetic version is to take a diffraction pattern with no bilateral symmetry, translate it by known amounts, run the RSS algorithm, and check whether the inferred translations match the imposed ones; a mismatch directly invalidates the alignment as a general-purpose correction.","tokens_in":8341,"feed_emoji":"🔬","tokens_out":3655,"duration_ms":42819,"temperature":0.7,"pith_summary":"This paper argues that the main obstacle to applying principal component analysis and k-means clustering to reflection high-energy electron diffraction (RHEED) videos is image drift: sample and stage motion translate the pattern across the screen, and the algorithms then learn the motion instead of the surface. The authors introduce a residual-sum-of-squares (RSS) alignment step that mirrors each RHEED frame to locate the specular spot and recenters the frames, along with intensity rescaling that enhances weak Kikuchi bands. With drift removed, PCA eigenvalues and k-means clusters from artificially drifted recordings match the originals, and the approach enables the first quantitative comparison of RHEED videos from two separately grown LaFeO3 film samples. If correct, this preprocessing pathway lets new growths be projected onto eigenvectors and centroids of past samples in real time, which is a step toward machine-learning-accelerated film synthesis.","feed_headline":"Mirror alignment cancels drift in RHEED videos for machine learning","feed_subtitle":"With pattern motion removed, PCA and k-means compare two LaFeO3 growths quantitatively for the first time.","key_machinery":"The central object is the residual sum of squares (RSS) alignment procedure. For each RHEED frame, the code mirrors the image across a vertical axis, slides the mirror in one-pixel steps, and takes the mirror position with minimum summed squared difference as the horizontal center of the pattern; a low-intensity filter that zeroes values below 90% of the maximum specular intensity keeps background signal from dominating the RSS calculation. A second pass, cropping to the specular spot and mirroring across a horizontal axis, fixes the vertical center. This turns frame-to-frame translation into a removable nuisance and is what allows PCA and k-means to be run on combined recordings from different samples, because the algorithm aligns both drift within a single growth and the systematic offset between separately mounted samples.","core_discovery":"On its own terms, the paper establishes that a mirror-symmetry-based RSS alignment algorithm removes the effect of pattern translation during RHEED recording, so that PCA and k-means cluster on physical surface features rather than on image motion. Applying this alignment to a recording with an artificial drift reproduces nearly identical eigenvalues, eigenvectors, and cluster centroids as the original recording, whereas the unaligned drifted recording produces components that encode the translation itself. Applying the same procedure across two LaFeO3 films grown under nominally identical conditions shows that both begin in the same substrate cluster and then diverge, with the second sample losing Kikuchi bands and sharp spots, indicating the onset of island formation or amorphization that would be obscured by unaligned PCA and clustering. This cross-sample comparison, which the paper identifies as a first for PCA and k-means on RHEED data, is what the RSS alignment makes possible.","pith_inferences":["Editorial inference: the success of RSS alignment depends on the pattern retaining bilateral symmetry; at off-symmetry azimuths, or during early nucleation when spots dominate, the mirrored-image minimum may not correspond to the true center, so the method should be validated against independently known translations at several azimuths before generalizing.","Editorial inference: the intensity transforms (power, inverse power) mainly aid visual interpretation and do not change the PCA results, which implies that a fully autonomous ML pipeline could skip the nonlinear rescaling and save computation, using only cropping and alignment.","Editorial inference: because RSS alignment centers to within a couple of pixels, the residual translation sets a floor on the smallest diffraction features the PCA and k-means can faithfully separate; features smaller than the residual jitter will be smeared into shared components and could be missed.","Editorial inference: extending the cross-sample comparison to arbitrary chamber geometries would require a more general feature-based or two-dimensional correlation alignment rather than mirror symmetry, since the symmetry assumption is tied to the specular geometry of the specific electron-beam and screen arrangement used here."],"forward_implications":["The preprocessing recipe of cropping, frame decimation, intensity transformation, and RSS alignment makes PCA and k-means results reproducible across separately recorded growths rather than only within one video.","New RHEED videos can be projected onto the eigenvector images and k-means centroid images of a library of previous samples, turning unsupervised clustering into a quantitative similarity check against known good and defective growths.","The RSS alignment approach should transfer to any RHEED geometry in which the diffraction pattern is mirror-symmetric, including other perovskite oxide films and other molecular beam epitaxy systems.","Eigenvalue time traces reveal oscillatory structure too weak for conventional specular-spot intensity monitoring, so the analysis extracts additional surface information from the same RHEED data without changing the growth conditions.","Because alignment works best for videos that begin with clean substrates and sharp features, it is most directly applicable to the early-growth regime, with performance expected to degrade as the pattern becomes faint or amorphous."],"supporting_citations":[{"why":"Supplies the LaFeO3 film growth procedure on Nb-doped SrTiO3 substrates that produced the RHEED videos analyzed here.","marker":"[21]"},{"why":"Provides the PCA, k-means, and video-capture workflow that this paper follows and extends with alignment and preprocessing.","marker":"[12]"},{"why":"Establishes the earlier big-data RHEED analysis framework with PCA and matrix factorization that this work builds on.","marker":"[11]"},{"why":"Reports comparative dimension-reduction characterization of thin-film growth, the context for cross-sample RHEED comparisons.","marker":"[14]"},{"why":"Defines the conventional RHEED intensity-oscillation analysis whose limitations motivate the machine-learning preprocessing approach.","marker":"[2]"},{"why":"Supplies background on RHEED from epitaxially grown thin films, grounding the claim that in situ pattern information reflects film evolution.","marker":"[1]"}],"fun_headline_variants":["Mirror alignment cancels drift in RHEED for ML","First cross-sample RHEED clustering via drift correction","Drift-free RHEED enables quantitative comparison of growths","RSS alignment lets PCA and k-means compare two LaFeO3 films"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole correction rides on the claim that a RHEED pattern is mirror-symmetric across its vertical axis and that the specular spot alone is symmetric about a horizontal axis; if a chosen azimuth, surface termination, shadow edge, or refraction effect breaks that symmetry, the RSS alignment will find the wrong center and reintroduce false features into the PCA and clusters.","fun_headline_variants_meta":{"raw":{"variants":["Mirror alignment cancels drift in RHEED for ML","First cross-sample RHEED clustering via drift correction","Drift-free RHEED enables quantitative comparison of growths","RSS alignment lets PCA and k-means compare two LaFeO3 films"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000722,"raw_usage":{"total_tokens":3282,"prompt_tokens":1031,"completion_tokens":2251,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":2178}},"tokens_in":647,"tokens_out":2251,"duration_ms":17069,"temperature":1.0,"reasoning_tokens":2178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:40:47.544802+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a RHEED video at an azimuth where the pattern is known to be asymmetric, apply RSS alignment, and compare the recovered center against the known true pattern location: if the recovered center shifts by more than a pixel when no motion is applied, or fails to track an imposed pixel offset, the symmetry premise fails. A simpler synthetic version is to take a diffraction pattern with no bilateral symmetry, translate it by known amounts, run the RSS algorithm, and check whether the inferred translations match the imposed ones; a mismatch directly invalidates the alignment as a general-purpose correction.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LaFeO3 film growth procedure on Nb-doped SrTiO3 substrates that produced the RHEED videos analyzed here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PCA, k-means, and video-capture workflow that this paper follows and extends with alignment and preprocessing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the earlier big-data RHEED analysis framework with PCA and matrix factorization that this work builds on."},{"cited_title":"Gliebe \\ and\\ author A","cited_arxiv_id":null,"evidence_quote":"Reports comparative dimension-reduction characterization of thin-film growth, the context for cross-sample RHEED comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the conventional RHEED intensity-oscillation analysis whose limitations motivate the machine-learning preprocessing approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies background on RHEED from epitaxially grown thin films, grounding the claim that in situ pattern information reflects film evolution."}],"review_version":1}