{"id":"88863fbc-1041-454a-942c-be19c45a90a5","arxiv_id":"2607.24467","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Physics-aware postprocessing lets a 2D ResNet trained on 15 CdTe structures recover crystallographic directions from rotating RHEED videos at near frame-resolution accuracy.","lead":"A 2D ResNet plus physics-aware postprocessing, trained on RHEED from only 15 CdTe growths, finds crystallographic directions within about half a degree. That removes a manual bottleneck in MBE and is a concrete step toward closed-loop thin-film growth.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline 96%-within-±0.5° figure rests on only 5 distinct holdout structures with 10× resampling, all grown in one chamber over 4 months under near-identical conditions; structure-level variance and domain shift — not cross-material transfer — are the real soft spot for the \"ready for CdTe\" part","rationale":"The reader identified the right family of concern — the physics-aware postprocessing premise (constant angular velocity, mod-180° circular statistics, two-pass KDE cuts at δ=10°/5°) is only stress-tested on CdTe holdout splits and two out-of-period videos — but aimed it mostly at the cross-material generalization claim in the abstract/§VII. I agree the cross-material claim is unsupported and should be tempered, but I judge the more load-bearing version of the concern to sit one level closer: even the CdTe deployment claim rests on an evaluation whose effective sample size is 5 structures, resampled 10× to n=382, drawn from a narrow temporal and process envelope (one chamber, 4 months, Ts 380–420°C). The reader's verdict of CONDITIONAL with conditions on shipping code, demonstrating live rotator control, and tempering cross-material language already covers the remedy; my stress-test sharpens *why* those conditions matter and adds one cheap, decisive check (LOSO per-structure intercept errors, computable today from the public Zenodo dataset) that would either solidify or puncture the pooled 96% figure. Hence UNCHANGED: the concern reinforces the reader's conditions rather than overturning the verdict. Credit where due: the paper's structure-wise splits, public data, honest label-noise discussion (0.33° floor), the live-feed playback tests, and the 12.5 FPS out-of-distribution video are genuine evidence — this is a careful applied-methods paper, and my concern is about the statistical reach of the headline numbers, not about the method's validity.","tokens_in":19908,"tokens_out":3117,"duration_ms":114150,"concrete_test":"Run leave-one-structure-out (LOSO) over all 20 structures with the 2D ResNet-50 + postprocessing pipeline: for each structure, train on the other 19, recover the intercept on every rotation of the held-out structure, and report the 20 per-structure mean intercept errors (not the pooled per-rotation density). If all 20 per-structure means fall within ±1.0° and ≥90% within ±0.5° — including the CdTe:In, Ts=380°C, and two-step-growth structures — the deployment claim is supported at the structure level. If even one or two structures exceed ±1.0°, the pooled 96% figure was masking structure-dependent failure modes and the claim needs qualifying. Complement with one prospective rotation set recorded after a deliberate perturbation (e.g., RPM changed from 1 to 2, or camera re-aimed) and check intercept recovery against a manually labeled zero.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim's deployment-relevant numbers (intercept 0.217±0.157°, 96% within ±0.5°, \"essentially all\" within ±1.0° for the 15/5 split) are computed over n=382 \"test samples,\" but n is inflated: n = (rotations per structure) × (20−p) × 10 resamples (§V.A.1, Fig. 3). With p=15, the underlying independent units are only 5 structures × 6–8 rotations each; rotations within a structure are highly correlated (same surface, same camera state, same holder shadows), and the 10× resampling reuses them. The bootstrap CIs in Fig. 5b are taken over evaluation runs (resampling partitions/seeds), so they quantify partition noise, not between-structure variability. If per-structure intercept error varies by more than the reported ±0.157°, the \"96%/essentially all\" probabilities do not estimate the reliability of the next genuinely new growth — they estimate a mixture dominated by whichever 5 structures happened to be held out.\n\nCompounding this, the entire dataset spans one chamber, one operator team, GaAs substrates, Ts ∈ [380, 420]°C (Table S1), and 4 months (Oct 2024–Jan 2025). The deployment claim (\"ready for closed-loop deployment in future CdTe growth experiments\") is precisely about leaving this envelope: heater-flake shadows grow, camera geometry drifts, RPM changes, doped vs undoped variants appear. The only out-of-envelope evidence is two supplementary videos (one 2 years before, one 1 year after the training period, one at 12.5 FPS vs the 18 FPS training data) — and each contributes a single intercept check, both starting near the 0° direction (314° and 340°), so neither exercises arbitrary intercepts or degraded-pattern regimes. That is encouraging but far too thin to certify prospective reliability. Notably, the 15/5 holdout pool contains the atypical structures (CdTe:In doped, Ts=380°C, two-step 410/350°C growth); whether those are systematically harder is invisible in the pooled density of Fig. 5c. This is not a circularity or soundness flaw — the pipeline is","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript presents a neural-vision pipeline for locating crystallographic azimuths in rotating-substrate RHEED videos acquired during CdTe/GaAs MBE. An ImageNet-initialized 2D ResNet-50 regresses the angle of each frame modulo 180°; a kinematic postprocessor then determines rotation direction, removes wrap-around and outlier residuals using circular statistics and KDE, and estimates the zero-angle intercept. Models are trained with structure-level train/validation/holdout partitions and compared with a 3D ResNet using four-frame clips. With 15 training and 5 holdout structures, the postprocessed 2D model is reported to reach mean intercept error 0.217±0.157°, 96% of samples within ±0.5°, and all within ±1°, while training roughly 20 times faster than the 3D benchmark. The authors additionally report 10-FPS CPU inference, two out-of-period video tests, public data, and claim readiness for closed-loop CdTe deployment and transfer to other materials.","tokens_in":20423,"tokens_out":7561,"duration_ms":319617,"significance":"If the reported structure-level performance holds, this is a practically useful demonstration of automated azimuth recovery from full-rotation RHEED under a realistic small-data constraint. Notable strengths are the publicly archived dataset, explicit label-noise floor, structure-level holdouts to reduce leakage, separate validation checkpointing, detailed training configurations, resolution and validation-size ablations, direct 2D/3D comparisons, supplementary out-of-period inference examples, and CPU inference timing. The physics-aware circular estimator uses acquisition kinematics rather than redefining the target, and the comparison quantifies its substantial benefit. The work is therefore a credible step toward MBE decision support, although the present evidence supports prospective deployment less strongly than the abstract and conclusion state.","major_comments":[{"comment":"§V.A.1 and Fig. 5: for p=15, n=382 is rotations per structure × 5 structures × 10 resamples. Rotations from the same structure are highly correlated, and resampling them does not create independent growths. The Fig. 5b bootstrap over evaluation runs therefore measures split/training-seed variability, not uncertainty for a genuinely new structure. Consequently, the headline 0.217±0.157°, 96%-within-±0.5°, and “essentially all” claims may not estimate next-growth reliability. Please report per-structure errors and a structure-clustered or hierarchical bootstrap (ideally including leave-one-structure-out results), and restate the probabilities at the structure level. Relatedly, §V.A.2 says results span “all possible structure-level partitions,” whereas Fig. 3 describes only 10 sampled partitions.","section":"§V.A.1–2, Fig. 5"},{"comment":"§III.A states that 5–10% of frames are duplicated or skipped, but §V.A.3 and Eqs. (1)–(5) assign angle by frame ordinal, αk=α0+kΔα, assuming a known constant increment. Such frame errors create phase mismatch in exactly the kinematic model on which the central postprocessor relies. Please either correct/index frames using timestamps or demonstrate robustness with synthetic 5–10% duplication/dropout, FPS and RPM jitter, and alternative δ and Δαwin values. The deployment section should also explain how 10-FPS CPU inference handles the nominal 18-FPS stream without silently breaking the frame-index-to-angle map.","section":"§III.A, §V.A.3–7, §VI.A"},{"comment":"The abstract and §VII claim a “fully trained” system “ready for closed-loop deployment,” while §VI.A says automatic stopping of substrate rotation is only planned. The quantitative out-of-period evidence consists of two recorded videos with individual starting angles, not a prospective live-growth trial or a distribution over new growths, RPM changes, camera drift, and long-term heater-shadow evolution. Likewise, transfer to other materials is argued but not tested outside CdTe/GaAs. Please either provide prospective closed-loop results with explicit success/failure criteria or temper these statements to real-time inference and a transfer-ready methodology.","section":"Abstract, §VI.A, §VII"},{"comment":"The central efficiency conclusion compares 2D* at 15/5 with 3D* at 19/1. These use different holdout compositions and apparently different error populations, so the quoted 0.217±0.157° versus 0.205±0.071° is not a paired comparison. Since “comparable accuracy with less data and ∼20× less training time” is a main claim, please evaluate both pipelines on the same held-out structures (with each model’s own training structures excluded), report paired per-structure differences, and define exactly what population each standard deviation summarizes.","section":"§VI.C, Table IV"}],"minor_comments":[{"comment":"Figures 5 and 7 use “Unnnormalized probability density”; presumably “Unnormalized” is intended.","section":"Figs. 5 and 7"},{"comment":"The preprocessing list runs (i)–(v), but augmentation is then numbered (vi) even though it occurs after dataset construction and before training. Please clarify which operations are dataset-level and which are training-only.","section":"§III.B"},{"comment":"In the 3D dataset definition, ν is called an angular shift but is also used as a frame index, and m=χν then mixes degrees and frame counts. Defining an integer frame stride separately from Δα=0.33° would avoid ambiguity.","section":"§V.B"},{"comment":"The video-inference discussion is duplicated. In the second version, the 340° example states “360−340=40,” although the expected remaining distance is 20°; the first version has the correct value.","section":"Supporting Information, RHEED videos inference"},{"comment":"The total row contains the malformed entry “133 490–”; please format the total frame count unambiguously.","section":"Table S1"},{"comment":"The reported mean intercept error (0.217°) is below the stated one-frame label floor (0.33°). This is plausible after averaging hundreds of residuals, but the manuscript should explain the sub-frame estimation and whether zero-angle labels were assigned independently for different rotations.","section":"§III.A and §VI.A"},{"comment":"The dataset DOI is welcome. Please also state whether training/postprocessing code, configuration files, and trained weights are available; if not, pseudocode for the complete two-pass KDE intercept procedure would improve reproducibility.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The archived dataset and unusually complete methodological detail are strengths. The recommendation is driven by the mismatch between the current uncertainty analysis and the deployment-level conclusions, not by a concern that the physics-aware estimator is circular. A structure-clustered reanalysis and more measured claims should substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is simple: with constant-rotation kinematics and modular circular cleanup, a plain 2D ResNet-50 trained on ~15 CdTe growths recovers the crystallographic intercept from a full rotation about as well as a heavier 3D ResNet on 19, at ~20× less train time. That fills a real gap the citations document—most prior RHEED ML either needs pre-aligned azimuths or does not output quantitative directions from rotating video. Structure-level repeated splits, bootstrap CIs, explicit ~0.33° label noise, resolution/validation ablations, and head-to-head 2D/3D with and without postprocessing are done cleanly. Data are on Zenodo; inference at 10 FPS on the acquisition CPU is credible for stopping rotation.\n\nWhat they do well is the physics postprocessing (sign of Δα via circular concentration, two-pass KDE outlier cuts, circular mean of window intercepts). Bare 2D MAE stays well above 1°; after postprocessing the intercept drops under the practical ±0.5–1° band for the 15/5 split. ImageNet init and on-the-fly ROI-style aug are sensible for small scientific sets. Writing is clear and the free parameters are at least listed.\n\nSoft spots are proportional, not fatal. The headline 96%-within-±0.5° / 0.217±0.157° numbers rest on five holdout structures × correlated rotations × 10× resampling; the bootstrap is over partitions, not independent growths. Everything is one chamber, one team, GaAs, Ts ~380–420 °C, four months. Two out-of-period videos (single intercepts near 0°) are encouraging but thin for “ready for closed-loop deployment.” Cross-material generality is a pipeline claim, not a result. Postprocessing thresholds (δ=10°/5°, 180° windows) are fixed, not swept. None of that breaks the CdTe comparison; it just means the deployment sentence should be tempered.\n\nWho it is for: MBE groups doing II–VI buffers or anyone building in-situ vision loops with scarce labeled RHEED. Math and citations look fine; no circularity. I would send it to referees—ask for code, live rotator demo, and softer multi-material language. Worth engaging if you care about growth automation; skip if you only want fundamental surface science.","headline":"Solid limited-data RHEED azimuth method for CdTe; the 2D+postprocessing result is real, the “ready for closed-loop / any material” framing is ahead of the evidence.","tokens_in":21484,"tokens_out":612,"would_cite":true,"duration_ms":19588,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A 2D network plus physics postprocessing finds CdTe crystallographic directions from RHEED using only 15 training structures.","keywords":["Molecular Beam Epitaxy","RHEED","Computer Vision","Machine Learning","Real-Time Feedback","CdTe","ResNet","crystallographic alignment"],"falsifier":"Run the trained 15-structure 2D pipeline on a fresh CdTe growth (or a different II-VI material) whose true zero azimuth is independently fixed by an expert; if the recovered intercept systematically exceeds the 1-degree tolerance or fails under ordinary RPM or ROI drift, the central claim fails.","tokens_in":21060,"feed_emoji":"🔬","tokens_out":915,"duration_ms":17817,"temperature":0.7,"pith_summary":"Manual RHEED inspection during MBE is slow and error-prone: growers must hunt for a few symmetric crystallographic azimuths inside long rotation videos. This paper shows that a standard 2D ResNet, trained on patterns from only about 15 CdTe buffers and followed by a postprocessing step that enforces constant rotation speed and circular 180-degree geometry, recovers the crystallographic direction to roughly 0.2 degrees mean intercept error on held-out structures. That accuracy matches a heavier 3D ResNet trained on more data, runs in real time on the acquisition PC, and is presented as ready for closed-loop CdTe growth. The practical payoff is a reusable pipeline that can automate alignment when large RHEED libraries do not exist.","feed_headline":"15 CdTe growths teach a network to lock RHEED azimuths","feed_subtitle":"Physics postprocessing lets a simple 2D model match heavier 3D nets and run live on the growth PC","key_machinery":"Physics-aware postprocessing after 2D angle regression: sign-of-slope selection by circular concentration, then two-pass KDE outlier rejection on modular residuals over 180-degree windows, yielding a robust intercept estimate of the unknown starting azimuth.","core_discovery":"With physics-aware postprocessing that uses constant angular velocity and modular-180 circular statistics, a 2D ResNet-50 trained on 15 CdTe structures recovers the crystallographic intercept from full-rotation RHEED with mean absolute error 0.217 plus or minus 0.157 degrees on holdout, lying within 0.5 degrees in 96 percent of cases and within 1 degree essentially always—comparable to a postprocessed 3D ResNet trained on 19 structures and far better than bare per-frame regression.","pith_inferences":["The same constant-velocity circular postprocessing could be tried on other in-situ diffraction streams (LEED, XRD rocking) where absolute angle labels are scarce.","If residual label noise from duplicated screen-grab frames is the true floor, cleaner hardware capture should push intercept error still closer to the 0.33-degree frame spacing.","Once azimuth locking is routine, the natural next closed-loop target is real-time growth-mode or stoichiometry feedback that re-uses the same aligned image stream."],"forward_implications":["The released CdTe model can be dropped into closed-loop control to stop substrate rotation on a chosen crystallographic direction during growth.","Laboratories with only 10–15 RHEED rotations of a new material can reuse the same 2D-plus-postprocessing recipe instead of collecting hundreds of structures.","Consistent azimuth-tagged frames become available for later automated growth-quality or anomaly models that currently require hand-selected key directions.","Training finishes in tens of minutes on one GPU and inference runs at 10 FPS on the acquisition CPU, removing the need for specialized hardware at deploy time."],"fun_headline_variants":["15 CdTe runs train 2D ResNet to lock RHEED azimuths","Physics postprocessing lets 2D net hit 0.22° RHEED error","Sparse RHEED data suffices for neural crystallographic lock","2D ResNet+physics matches 3D on 15-structure CdTe set","Limited CdTe RHEED trains live MBE azimuth alignment"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that known constant rotation speed plus modular circular cleaning will keep turning noisy single-frame angle guesses into a trustworthy intercept on new growths and, by extension, on other materials that only have small RHEED libraries.","fun_headline_variants_meta":{"raw":{"variants":["15 CdTe runs train 2D ResNet to lock RHEED azimuths","Physics postprocessing lets 2D net hit 0.22° RHEED error","Sparse RHEED data suffices for neural crystallographic lock","2D ResNet+physics matches 3D on 15-structure CdTe set","Limited CdTe RHEED trains live MBE azimuth alignment"]},"model":"grok-4.5","effort":"low","cost_usd":0.005144,"raw_usage":{"total_tokens":1370,"prompt_tokens":769,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":51440000,"prompt_tokens_details":{"text_tokens":769,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":512,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":769,"tokens_out":89,"duration_ms":15882,"temperature":1.0,"reasoning_tokens":512,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T14:01:55.400884+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the trained 15-structure 2D pipeline on a fresh CdTe growth (or a different II-VI material) whose true zero azimuth is independently fixed by an expert; if the recovered intercept systematically exceeds the 1-degree tolerance or fails under ordinary RPM or ROI drift, the central claim fails.","supporting_citations":[],"review_version":1}