Pith. sign in

REVIEW 4 major objections 7 minor 53 references

Neural RHEED alignment with limited training data during CdTe MBE growth

T0 review · 4 major / 7 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A 2D network plus physics postprocessing finds CdTe crystallographic directions from RHEED using only 15 training structures.

desk verdict 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. read the letter →

arxiv 2607.24467 v1 pith:YMEIF32Y submitted 2026-07-27 cond-mat.mtrl-sci

Bart{l}omiej Turowski , Jakub J. Meixner , Róża Dziewiątkowska , Wojciech Zaleszczyk , Tomasz Wojciechowski , Valentine V. Volobuev , Marcin M. Wysokiński , Tomasz Wojtowicz This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords MolecularBeamEpitaxyRHEEDComputerVisionMachineLearningReal-TimeFeedbackCdTeResNetcrystallographicalignment
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

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

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

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

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.

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 (4)
  1. [§V.A.1–2, Fig. 5] §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.
  2. [§III.A, §V.A.3–7, §VI.A] §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.
  3. [Abstract, §VI.A, §VII] 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.
  4. [§VI.C, Table IV] 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.
minor comments (7)
  1. [Figs. 5 and 7] Figures 5 and 7 use “Unnnormalized probability density”; presumably “Unnormalized” is intended.
  2. [§III.B] 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.
  3. [§V.B] 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.
  4. [Supporting Information, RHEED videos inference] 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.
  5. [Table S1] The total row contains the malformed entry “133 490–”; please format the total frame count unambiguously.
  6. [§III.A and §VI.A] 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.
  7. [Data Availability] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: supervised holdout evaluation with kinematics-based postprocessing that does not redefine the target

full rationale

This is an empirical computer-vision paper. Crystallographic labels are assigned by human operators identifying the zero-angle direction independently of the network. The 2D/3D ResNets are trained by MSE regression on those labels and evaluated on structure-level holdout splits (repeated train-test splits, validation checkpoints by MAE). The physics-aware postprocessing (constant known Δα from substrate rotation, modular-180° residuals, circular mean, two-pass KDE outlier cuts) uses the acquisition kinematics stated in §V.A.3–7 to estimate the intercept α̂₀; it does not fit α₀ from the same quantity it then claims to predict, nor does it import a uniqueness theorem or load-bearing self-citation. Performance figures (intercept 0.217±0.157°, 96% within ±0.5° on the 15/5 split) are empirical holdout metrics, not identities by construction. Domain-shift and n-inflation concerns affect generalization risk, not circularity of the derivation chain. No steps reduce Eq. X to Eq. Y by definition or rename a fitted input as a prediction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

Load-bearing content is standard CV backbones plus domain kinematics of MBE RHEED acquisition. No new physical entities. Performance claims rest on labeled operator angles, constant RPM, 180° pattern similarity, ImageNet transfer, and hand-chosen postprocessing thresholds—not on a parameter-free theory of diffraction.

free parameters (5)
  • Outlier rejection thresholds δ (10° then 5°) and window Δα_win=180° = δ=10° then 5°; Δα_win default 180°
    Chosen for iterative KDE cleaning of modular residuals; directly control which frames enter the intercept estimate.
  • 2D training hyperparameters (lr=1e-3, wd=1e-4, 150 epochs, batch 2048, cosine+warmup) = AdamW defaults as in Table II
    Fixed without systematic search; affect achieved MAE before postprocessing.
  • On-the-fly augmentation ranges = as in Fig. 2 / §III.B
    Rotation ±5°, translation ±15%, scale ±7%, brightness/contrast ±15%, Gaussian σ=0.05—hand-set to emulate ROI placement noise.
  • 3D clip design (χ=3, ν=0.33°, m=last frame) and 3D LR schedule = 4-frame clips; Cosine Warm Restarts T=50, etc.
    Benchmark architecture hyperparameters fixed or lightly swept; change bare 3D MAE.
  • Intensity exclusion gates (mean<0.02, RMS contrast<0.01) and fixed ROI crop fractions = crop top 23%/left 13%/right 10%/bottom 30%
    Dataset-level filters and crop percentages define what the network sees.
assumptions (6)
  • domain assumption Substrate angular velocity is constant and known during each labeled rotation, so α_k = α_0 + k Δα.
    Stated in §V.A step 3; entire intercept estimator depends on it.
  • domain assumption RHEED patterns at azimuths differing by 180° are similar enough that labels may be folded into [0°,180°).
    §V.A dataset construction; drives modular arithmetic and wrap-around handling.
  • domain assumption ImageNet-pretrained ResNet-50 features are a useful initialization for grayscale RHEED regression despite domain gap.
    §V.A; authors report large gain vs scratch (MAE 3.1 vs 7.7) and adopt it throughout.
  • domain assumption Operator-identified zero-angle labels are accurate to roughly one frame (~0.33°) aside from grabber drop/dup noise.
    §III.A; sets the noise floor against which reported errors are interpreted.
  • standard math Standard ResNet residual convolutional architectures and MSE/MAE regression training are appropriate off-the-shelf tools.
    §V cites He et al. and common practice; no new architecture theory.
  • ad hoc to paper Holdout structures drawn from the same CdTe/GaAs MBE process distribution represent deployment conditions.
    Generalization and ‘ready for closed-loop’ claims assume future growths resemble the 20-structure corpus (Oct 2024–Jan 2025).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural RHEED alignment with limited training data during CdTe MBE growth." pith.science (2026). https://pith.science/paper/YMEIF32Y

@misc{pith2026260724467,
  author       = {Pith},
  title        = {Pith review of: Neural RHEED alignment with limited training data during CdTe MBE growth},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YMEIF32Y}},
  note         = {Machine review of arXiv:2607.24467}
}
read the original abstract

We introduce a data-efficient neural-vision assisted method to automate crystallographic alignment during molecular beam epitaxy (MBE) growth. Trained on reflection high-energy electron diffraction (RHEED) patterns from only 15 CdTe structures, our model - enabled by physics-aware postprocessing - reliably infers crystallographic directions, replacing manual frame-by-frame inspection. To this end, we design, test, and critically compare neural-network architectures based on 2D and 3D ResNet configurations, both with and without postprocessing that leverages the physical constraints of RHEED image acquisition. Our work delivers (i) a fully trained neural system ready for closed-loop deployment in future CdTe growth experiments and (ii) a generalizable pipeline for new materials where access to diverse RHEED datasets is limited. More broadly, this study represents a step toward AI-driven MBE growth and demonstrates the potential of machine-learning-assisted automation in thin-film synthesis.

Figures

Figures reproduced from arXiv: 2607.24467 by the authors.

Figure 1
Figure 1. FIG. 1: Schematic overview of the dataset-preparation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Schematic flow illustrating the on-the-fly [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Visualization of the division of images from the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Visualization of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: FIG. 7: Unnormalized probability density of the MAE [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 1 linked inside Pith

  1. [1]

    V. V. Volobuev, P. S. Mandal, M. Galicka, O. Caha, J. Sánchez-Barriga, D. Di Sante, A. Varykhalov, A. Khiar, S. Picozzi, G. Bauer, P. Kacman, R. Buczko, O. Rader, and G. Springholz, Advanced Materials29, 1604185 (2017)

    2017

  2. [2]

    Kazakov, V

    A. Kazakov, V. V. Volobuev, C.-W. Cho, B. A. Piot, Z. Adamus, T. Wojciechowski, T. Wojtowicz, G. Springholz, and T. Dietl, Phys. Rev. B111, 245419 (2025)

    2025

  3. [3]

    Polaczyński, G

    J. Polaczyński, G. Krizman, A. Kazakov, B. Tur- owski, J. B. Ortiz, R. Rudniewski, T. Wojciechowski, P. Dłużewski, M. Aleszkiewicz, W. Zaleszczyk, B. Kurowska, Z. Muhammad, M. Rosmus, N. Ol- szowska, L.-A. de Vaulchier, Y. Guldner, T. Wojtowicz, and V. V. Volobuev, Materials Today75, 135 (2024)

    2024

  4. [4]

    Carmody, A

    M. Carmody, A. Yulius, D. Edwall, D. Lee, E. Pi- quette, R. Jacobs, D. Benson, A. Stoltz, J. Markunas, A. Almeida, and J. Arias, Journal of Electronic Materi- als41, 2719 (2012)

    2012

  5. [5]

    W.Pan, R.J.Gu, Z.K.Zhang, W.Lei, G.A

    W. W.Pan, R.J.Gu, Z.K.Zhang, W.Lei, G.A. Umana- Membreno, D. J. Smith, J. Antoszewski, and L. Faraone, Journal of Electronic Materials51, 4869 (2022)

    2022

  6. [6]

    Bansal, A

    S. Bansal, A. Jain, S. Kumar, A. Kumar, P. R. Kumar, K. Prakash, M. S. Soliman, M. S. Islam, and M. T. Islam, Scientific Reports14, 28230 (2024)

    2024

  7. [7]

    C. Shen, W. Zhan, M. Li, Z. Sun, J. Tang, Z. Wu, C. Xu, B. Xu, C. Zhao, and Z. Wang, Journal of Semiconductors 45, 031301 (2024)

    2024

  8. [8]

    Daniluk, B

    A. Daniluk, B. Daniluk, and G. M. Wójcik, Computer Physics Communications308, 109467 (2025)

    2025

Show all 53 references
  1. [9]

    Daniluk, A

    B. Daniluk, A. Daniluk, and G. M. Wójcik, Computer Physics Communications325, 110186 (2026)

    2026

  2. [10]

    Kaufmann, C

    K. Kaufmann, C. Zhu, A. S. Rosengarten, D. Maryanovsky, T. J. Harrington, E. Marin, and K. S. Vecchio, Science367, 564 (2020)

    2020

  3. [11]

    Xiong, C

    G. Xiong, C. Wang, Y. Yan, L. Zhang, and Y. Su, Com- putational Materials Science233, 112718 (2024)

    2024

  4. [12]

    P. M. Vecsei, K. Choo, J. Chang, and T. Neupert, Phys- ical Review B99, 245120 (2019)

    2019

  5. [13]

    Ziatdinov, O

    M. Ziatdinov, O. Dyck, A. Maksov, X. Li, X. Sang, K. Xiao, R. R. Unocic, R. Vasudevan, S. Jesse, and S. V. Kalinin, ACS Nano11, 12742 (2017)

    2017

  6. [14]

    Petkovic, L

    M. Petkovic, L. Vieira, and N. Dropka, Progress in Crys- tal Growth and Characterization of Materials71, 100689 (2025)

    2025

  7. [15]

    R. K. Vasudevan, A. Tselev, A. P. Baddorf, and S. V. Kalinin, ACS Nano8, 10899 (2014)

    2014

  8. [16]

    S. R. Provence, S. Thapa, R. Paudel, T. K. Truttmann, A. Prakash, B. Jalan, and R. B. Comes, Physical Review Materials4, 083807 (2020)

    2020

  9. [17]

    P. T. Gemperline, R. Paudel, R. K. Vasudevan, and R. B. Comes, Journal of Vacuum Science & Technology A43, 032701 (2025)

    2025

  10. [18]

    Kwoen and Y

    J. Kwoen and Y. Arakawa, Crystal Growth & Design20, 5289 (2020)

    2020

  11. [19]

    Kwoen and Y

    J. Kwoen and Y. Arakawa, Journal of Crystal Growth 593, 126780 (2022)

    2022

  12. [20]

    Muetzel, V

    N. Muetzel, V. Luu, S. Bey, M. Abdul Karim, K. Yoshimura, X. Liu, M. Gebran, and B. A. Assaf, AIP Advances16, 035317 (2026)

    2026

  13. [21]

    J. R. Chin, I. A. Moses, M. Wang, M. B. Frye, M. Yu, N. Nayir, M. Hilse, A. C. van Duin, S. Law, W. Reinhart, and L. M. Garten, Materials Today Advances28, 100640 (2025)

    2025

  14. [22]

    Liang, V

    H. Liang, V. Stanev, A. G. Kusne, Y. Tsukahara, K. Ito, R. Takahashi, M. Lippmaa, and I. Takeuchi, Physical Re- view Materials6, 063805 (2022)

    2022

  15. [23]

    Yoshinari, Y

    A. Yoshinari, Y. Ando, T. Matsumura, M. Kotsugi, S. Hasegawa, R. Hobara, and N. Nagamura, Science and Technology of Advanced Materials: Methods6, 2666924 (2026)

    2026

  16. [24]

    Chong, T

    M. Chong, T. G. Rhee, Y. G. Khim, M.-H. Jung, Y.-M. Kim, H. Y. Jeong, H.-S. Kim, Y. J. Chang, and H. J. Kim, Applied Surface Science669, 160547 (2024)

    2024

  17. [25]

    H. J. Kim, M. Chong, T. G. Rhee, Y. G. Khim, M.-H. Jung, Y.-M. Kim, H. Y. Jeong, B. K. Choi, and Y. J. Chang, Nano Convergence10, 10 (2023)

    2023

  18. [26]

    T. C. Kaspar, S. Akers, H. W. Sprueill, A. H. Ter- Petrosyan, J.A.Bilbrey, D.Hopkins, A.Harilal, J.Chris- tudasjustus, P. Gemperline, and R. B. Comes, Journal of Vacuum Science & Technology A43, 032702 (2025)

    2025

  19. [27]

    Y. Guo, P. Meisenheimer, S. Qin, X. Zhang, J. Goddy, R. Ramesh, L. W. Martin, and J. Agar, ACS Applied Materials & Interfaces17, 24485 (2025)

    2025

  20. [28]

    C.Shen, W.Zhan, K.Xin, M.Li, Z.Sun, H.Cong, C.Xu, J. Tang, Z. Wu, B. Xu, Z. Wei, C. Xue, C. Zhao, and Z. Wang, Nature Communications15, 2724 (2024)

    2024

  21. [29]

    Anjum, H.-Y

    S. Anjum, H.-Y. Lee, and H.-K. Noh, Crystal Growth & Design23, 4138 (2023)

    2023

  22. [30]

    Khaireh-Walieh, A

    A. Khaireh-Walieh, A. Arnoult, S. Plissard, and P. R. Wiecha, Crystal Growth & Design23, 892 (2023)

    2023

  23. [31]

    C. Shen, W. Zhan, J. Tang, Z. Wu, B. Xu, C. Zhao, and Z. Wang, ACS Applied Materials & Interfaces16, 18213 (2024), pMID: 38554077

    2024

  24. [32]

    Khaireh-Walieh, A

    A. Khaireh-Walieh, A. Arnoult, S. Plissard, and P. R. Wiecha, Crystal Growth & Design25, 7438 (2025)

    2025

  25. [33]

    M. Yu, I. A. Moses, W. F. Reinhart, and S. Law, ACS Applied Materials & Interfaces17, 34707 (2025), pMID: 40434265

    2025

  26. [34]

    Ohkubo, Z

    I. Ohkubo, Z. Hou, J. Lee, T. Aizawa, M. Lippmaa, T. Chikyow, K. Tsuda, and T. Mori, Materials Today Physics16, 100296 (2021)

    2021

  27. [35]

    Y. K. Wakabayashi, T. Otsuka, Y. Krockenberger, H. Sawada, Y. Taniyasu, and H. Yamamoto, APL Ma- terials7, 101114 (2019)

    2019

  28. [36]

    S. B. Harris, A. Biswas, S. J. Yun, K. M. Roccapri- ore, C. M. Rouleau, A. A. Puretzky, R. K. Vasudevan, 13 D. B. Geohegan, and K. Xiao, Small Methods8, 2301763 (2024)

    2024

  29. [37]

    Y. B. Zheng, C. Blake, L. Mravac, F. Zhang, Y. Chen, and S. Yang, npj Computational Materials11, 327 (2025)

    2025

  30. [38]

    veeco.com/products/genxplor-rd-mbe-system/

    Veeco GENxplor R&D MBE System,https://www. veeco.com/products/genxplor-rd-mbe-system/

  31. [39]

    Veeco - Molly 2000,https://www.veeco.com/products/ growth-control-scheduling-software-for-mbe-systems/

    2000

  32. [40]

    Turowski, J

    B. Turowski, J. J. Meixner, R. Dziewiątkowska, W. Za- leszczyk, T. Wojciechowski, V. V. Volobuev, M. M. Wysokiński, and T. Wojtowicz, Supporting information

  33. [41]

    Keys, IEEE Transactions on Acoustics, Speech, and Signal Processing29, 1153 (1981)

    R. Keys, IEEE Transactions on Acoustics, Speech, and Signal Processing29, 1153 (1981)

    1981

  34. [42]

    K. He, X. Zhang, S. Ren, and J. Sun, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 770 (2016)

    2016

  35. [43]

    S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He, Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 5987 (2017)

    2017

  36. [44]

    Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 11976 (2022)

    2022

  37. [45]

    Tan and Q

    M. Tan and Q. V. Le, Proceedings of the 38th Interna- tional Conference on Machine Learning (ICML) , 10096 (2021)

    2021

  38. [46]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, Inter- nationalConferenceonLearningRepresentations(ICLR) (2021)

    2021

  39. [47]

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 10012 (2021)

    2021

  40. [48]

    Touvron, M

    H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, and H. Jégou, Proceedings of the 38th International Confer- ence on Machine Learning (ICML) , 10347 (2021)

    2021

  41. [49]

    A. Y.-T. Wang, R. J. Murdock, S. K. Kauwe, A. O. Oliynyk, A. Gurlo, J. Brgoch, K. A. Persson, and T. D. Sparks, Chemistry of Materials32, 4954 (2020)

    2020

  42. [50]

    Stuckner, B

    J. Stuckner, B. Harder, and T. M. Smith, npj Computa- tional Materials8, 200 (2022)

    2022

  43. [51]

    C. C. Price, Y. Li, G. Zhou, R. Younas, S. S. Zeng, T. H. Scanlon, J. M. Munro, and C. L. Hinkle, Nano Letters 24, 14862 (2024)

    2024

  44. [52]

    Loshchilov and F

    I. Loshchilov and F. Hutter, Decoupled weight decay reg- ularization (2019), arXiv:1711.05101 [cs.LG]

    2019 arXiv

  45. [53]

    Raw angle

    S. B. Harris, C. M. Rouleau, K. Xiao, and R. K. Vasude- van, npj Computational Materials10, 105 (2024). Supporting Information Neural RHEED alignment with limited training data during CdTe MBE growth Bartłomiej Turowski1, Jakub J. Meixner1, Róża Dziewiątkowska1, Wojciech Zales...

    2024

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.