Pith. sign in

REVIEW 3 major objections 6 minor 59 references

Cross-Modal Characterization of Thin Film MoS$_2$ Using Generative Models

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that Raman spectra can predict a UMAP-based morphology representation of thin-film MoS$_2$ with test $R^2$ up to 0.69, and that autoencoder models generate full Raman and photoluminescence spectra from other modalities…

desk verdict A useful, honest feasibility study with released code and data; the headline Raman-to-AFM claim needs a UMAP-protocol fix and corrected confidence intervals before I would trust the numbers. read the letter →

arxiv 2505.24065 v2 pith:FRDPL4DB submitted 2025-05-29 cond-mat.mtrl-sci cs.LGphysics.app-ph

Isaiah A. Moses , Chen Chen , Joan M. Redwing , Wesley F. Reinhart This is my paper
classification cond-mat.mtrl-scics.LGphysics.app-ph
keywords cross-modallearningRamanspectroscopyphotoluminescenceatomicforcemicroscopythinfilmMoS2autoencoderUMAPmachinecharacterization
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

This paper tries to show that complementary materials-characterization techniques can stand in for one another in thin-film MoS$_2$. Using a library of atomic force microscopy (AFM) images, Raman spectra, and photoluminescence (PL) spectra from metal-organic chemical vapor deposition growths, the authors claim that Raman spectra can predict the AFM-derived latent morphology representation with test $R^2$ up to 0.69, and that autoencoder models can generate full Raman and PL spectra from other modalities with test Pearson $R$ around 0.89–0.96. If true, this would let researchers estimate slow, instrument-heavy microscopy morphology from faster spectroscopy, and synthesize missing spectral data, accelerating growth-parameter optimization.

What carries the argument

The load-bearing machinery is the unsupervised UMAP embedding of AFM images. A pretrained deep residual network extracts 2048 features per image; UMAP projects them to two dimensions whose principal axes are interpretable as feature size and structural order, based on correlations with roughness, grain density, and image dissimilarity. That embedding becomes the regression target for Raman spectra, and the same image features feed autoencoder decoders that generate whole spectra. The autoencoders use binary cross-entropy loss with normalized intensities and fully connected decoder heads, which the paper finds generalize better than recurrent or transformer decoders on this dataset.

What would settle it

Re-run the pipeline with UMAP fitted only on training AFM images, then transform test images through that fixed embedding and retrain the Raman regressors; if the held-out $R^2$ for UMAP-1 falls well below 0.69, the deployable claim is weakened. A complementary check is to split the test set by growth batch and see whether the Raman-to-embedding correlation persists across unseen synthesis conditions.

Watch

Extended reading notes

Core claim

The central claim is that the two axes of a UMAP embedding of AFM height images are learnable from Raman spectra, even though directly predicting hand-chosen image features such as roughness or grain density from Raman generally fails. The paper reports that the UMAP-1 axis, which correlates with structural order and surface smoothness, is predicted with test $R^2$ up to 0.69 using support vector regression, while UMAP-0, related to feature size, reaches 0.59 with Ridge regression. The same framework then shows that autoencoders with fully connected decoder heads can reconstruct complete Raman and PL spectra from other modalities, with held-out test Pearson correlations of 0.89 to 0.96, and that fusing AFM with one spectrum yields full spectra of the other with comparable accuracy.

Load-bearing premise

The load-bearing premise is that the UMAP embedding of the AFM images is a stable, physically meaningful representation; the paper does not state whether UMAP is fitted on the training split only, so if it used all images the reported test $R^2$ may overstate deployable accuracy.

Editorial extensions

If this is right

  • Rapid spectroscopy could screen growth quality by estimating AFM-derived morphology coordinates without waiting for AFM scans.
  • Full Raman or PL spectra can be synthesized from the other modality, making it possible to fill missing characterization channels in historical or sparse datasets.
  • Fusing AFM with one spectrum produces the other spectrum with test correlation comparable to single-modality inputs, suggesting multimodal fusion is a practical route.
  • The interpretable UMAP axes give a low-dimensional morphology descriptor that machine learning can predict better than individual physical image features.
  • The workflow transfers in principle to other materials systems and larger datasets, where the authors expect improved generalization.

Reading between the lines

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

  • A deployable version of the regression pipeline would need to fit UMAP on training images only and transform held-out images afterwards; if the reported embedding was fit on all images, the test $R^2$ could overstate real-world performance on a new Raman spectrum.
  • The high whole-spectrum correlations should not be read as high accuracy on quantitative peak features: the paper's $R^2$ values for predicting the A$_{1g}$–E$_{2g}^1$ split and FWHM from generated spectra are 0.52 or lower on held-out tests, so generated spectra may be better for pattern-level synthesis than for precise metrology.
  • The same approach could be tested as a transfer-learning baseline for other two-dimensional materials, where Raman peak positions shift with composition and layer number, by fine-tuning the encoder on a small target-material dataset.
  • A practical closed-loop growth optimizer could use the Raman-to-UMAP regressor as a fast surrogate reward, flagging samples whose predicted morphology coordinate falls in a desirable region before committing to slower AFM characterization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript investigates whether Raman and PL spectroscopy can be used to infer AFM-derived morphological information in MOCVD-grown monolayer MoS2. A ResNet152 + UMAP embedding of AFM images is used as a low-dimensional target; regression models (MLP, Ridge, SVR) are trained to predict these latent coordinates from Raman and PL spectra. Autoencoder-based generative models are then trained to reconstruct full Raman and PL spectra from the other modalities and from AFM images, including fusion models that combine AFM with one spectral modality to generate the other. The authors report test R2 values up to 0.69 for Raman-to-UMAP-1 prediction and Pearson correlations above 0.88 for generated spectra on held-out test sets. They conclude that cross-modal ML can accelerate characterization and guide growth optimization. The paper provides open code and data for reproducibility.

Significance. If the central claims are correct, the work is a valuable demonstration of cross-modal characterization in a realistic materials dataset with grouped train/test splits, and the open data/code release is a notable strength. The finding that Raman spectra carry information about AFM-derived morphology in a latent form, and that full spectra can be generated across modalities, could inform high-throughput characterization workflows. However, the quantitative reliability of the main regression result and its out-of-sample validity are not yet fully established because of the issues detailed below. The core idea is sound and potentially impactful, but several load-bearing technical points need to be corrected or clarified before the claims can be accepted.

major comments (3)
  1. [Table 3, §2.4] The bootstrap confidence intervals reported in Table 3 are internally inconsistent with the test R2 values for Ridge and for SVR on UMAP-0. Specifically, Ridge UMAP-0 test R2 = 0.59 with 95% CI [0.92, 1.00], Ridge UMAP-1 test R2 = 0.68 with CI [0.93, 1.00], and SVR UMAP-0 test R2 = 0.56 with CI [0.66, 0.83]; none of these intervals contain the corresponding test R2. As described in §2.4, the CIs are obtained by bootstrap resampling the training set and evaluating the models on the test set; it is statistically impossible for a 95% CI to exclude the point estimate it is built around. This suggests the CI was computed for a different quantity (for example, training R2) or there is a labeling error. Because the paper cites these intervals to claim that Ridge has 'very high precision' and to compare model reliability, this error is load-bearing and must be corrected.
  2. [§2.3, §3.2] The paper does not state whether the UMAP embedding is fitted on the training/validation split only or on the full set of AFM images before the grouped split described in Table 1. The text in §2.3 says the number of neighbors was 'set to half the sample size (193)', and since the AFM-Raman dataset contains 370 total samples (306 train/validation + 64 test), 193 is close to half of the full set, which suggests the UMAP was fitted using all images. If this is the case, the test images contribute to the construction of the embedding manifold, so the test UMAP coordinates are not truly out-of-sample: they are jointly optimized with the training coordinates and would not be obtained at inference time, when a new Raman spectrum would need to be mapped to an embedding produced by a UMAP fitted on training data only. Because the central claim of Raman-to-AFM-latent generalization (test R2 up to 0.69) depends on the test targets being genuine held-out coordinates, the authors must specify the exact protocol (for example, fitting UMAP on the train+validation split and transforming the test ResNet features with that fitted UMAP) and, if the current analysis used a full-data fit, re-evaluate the regressions under a train-only embedding. The authors' own statement in §2.3 that the manifold depends on 'which data subsets are used' makes this protocol detail essential.
  3. [§3.2, Table S3] The direct regression of physically meaningful AFM features from Raman spectra in Table S3 yields mostly near-zero or even negative test R2 values (for example, step edges: -0.07 for Ridge; grain density: 0.44 best but -0.90 for SVR; RMS roughness: 0.33; DoG blob: 0.48; dissimilarity: 0.54). Yet the paper argues in §3.1 and §3.2 that the UMAP latent dimensions are physically meaningful because of their correlations with these same features (Table 2), and that Raman-predicted UMAP-1 therefore represents a genuine physical relationship. If UMAP-1 is strongly correlated with RMS roughness (R = -0.74) and dissimilarity (R = -0.77), it is not obvious why a model that can predict UMAP-1 with test R2 = 0.68 cannot predict a significant fraction of variance in those direct features. The paper's explanation that latent features capture broader properties may be correct, but this tension needs a quantitative reconciliation, for example by showing that the predicted UMAP-1 corresponds to a nonlinear combination of physical features that is more learnable than any individual feature, or by reporting the R2 for a linear projection of physical features onto the predicted latent. Without such a check, the claim that the regression has 'learned a genuine physical relationship' is not fully supported.
minor comments (6)
  1. [§2.3] The phrase 'half the sample size (193)' is ambiguous because the total number of AFM images used for UMAP is not explicitly stated at that point; please specify the exact count and whether 193 corresponds to the full dataset or to the training subset.
  2. [Table 3] The column grouping 'test 95% CI' spans MLP, Ridge, and SVR, but the intervals are shown in three sub-columns aligned under each model; consider separating the CI from the R2 table or using explicit subheadings to avoid ambiguity.
  3. [Supporting Information, Figures S2 and S3] The captions for Figures S2 and S3 contain a typo: 'Samaple' should be 'Sample'.
  4. [§2.5] The sentence 'the number of nodes in each layer et to be half of the preceding layer' should read 'is set to half of the preceding layer'.
  5. [§4] The conclusion describes the generative models as 'self-supervised'; since the cross-modal autoencoders are trained with explicit target spectra, the term 'supervised' or 'cross-modal reconstruction' would be more precise.
  6. [Table S2] In the spectra-spectra autoencoder row, the hidden dimensions are given as lists such as '214, 840, 814'; it would be helpful to state explicitly whether these correspond to the three fully connected layers in the encoder, decoder, or both.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Raman-to-AFM-latent and cross-spectral generation are evaluated against independent held-out targets; self-citations are not load-bearing.

full rationale

The central derivation chain is self-contained. The regression targets are UMAP/PCA coordinates obtained from AFM images through a fixed ResNet152 feature extractor plus UMAP, while the inputs are Raman spectra; no target value is defined in terms of the input spectrum, and the grouped train/test split (Section 2.2) keeps sample identities out of the training fit. Similarly, the generative models map one measured modality to another (AFM to Raman/PL, Raman to PL and vice versa) and are scored by comparing generated spectra to held-out measured spectra (Tables 4-5), so the outputs are not contained in the inputs. The few self-citations (Refs. 42 and 50) support method choices (UMAP suitability and ResNet transfer performance) but are not the evidence for the cross-modal generalization, which rests on held-out R2 and Pearson R values. Two caveats should be noted as correctness risks rather than circularity. First, Section 2.3 does not state that the UMAP embedding was fit on the training split only; the n_neighbors=193 statement and the remark that the manifold depends on 'which data subsets are used' leave open the possibility that test images participated in constructing the embedding, which would make the test latent coordinates not strictly out-of-sample and could inflate the reported test R2. This is a potential data-leakage/evaluation-protocol issue, not a definitional reduction of the Raman prediction to its input. Second, the passing mention of autoencoders that 'directly predict the A1g-E1 2g split and FWHM from the Raman and PL' (Section 2.5) is tautological because those quantities are extracted from the same spectra, but no central claim or reported result depends on that auxiliary model. Overall, the paper's core cross-modal predictions are independent of their inputs by construction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claims rest on tuned hyperparameters (UMAP parameters, regression and autoencoder architectures) and on domain assumptions that ImageNet features transfer to AFM images and that the cleaned dataset is representative. These are standard ML practice but remain assumptions the paper does not independently validate. No new physical entities are introduced.

free parameters (7)
  • UMAP n_neighbors = 193
    Hand-selected as half the sample size to favor global structure; directly controls the embedding that becomes the regression target (Section 2.3).
  • UMAP random_state = fixed integer (not specified)
    Set to ensure reproducibility; different seeds produce different embeddings, affecting all downstream results (Section 2.3).
  • MLP hyperparameters = 5 hidden layers, sizes 71/57/48/48/34, dropout 0.288/0.162/0.129/0.183/0.284, L1 2.45e-4, lr 6.45e-3, batch 62
    Tuned with Bayesian optimization (Table S1); the reported generalization is conditional on these choices.
  • Ridge alpha = 10.0
    Tuned by grid search (Table S1); alpha is a regularization strength that affects test R2.
  • SVR hyperparameters = degree 2, C varies (10, 10, 100, 50), gamma scale, RBF kernel
    Tuned by grid search (Table S1); C and gamma strongly influence the fit.
  • Autoencoder hyperparameters = e.g., PL-to-Raman AE hidden dims 214/840/814, dropout 0.4117, lr 2.076e-3, batch 32; other models use their own…
    Optimized via Bayesian optimization with 10-fold CV (Table S2); the generative performance depends on these choices.
  • Data filters = 1 um scan size, 4 mW laser power, 202-point Raman range
    Hand-chosen cleaning rules in Section 2.2 that reduce the data to about 30%, excluding measurements that might change peak shapes.
assumptions (4)
  • domain assumption ImageNet-pretrained ResNet152 feature extractor transfers effectively to AFM height images.
    Section 2.3 uses the average pool layer of ResNet152 trained on ImageNet to represent AFM images without fine-tuning on material images; the paper relies on this transfer without validation against a domain-trained encoder.
  • domain assumption UMAP with n_neighbors=193 and cosine distance preserves the global morphological structure of the AFM data.
    Section 2.3 selects these parameters by hand; the paper acknowledges that the learned manifold depends on hyperparameters and data subsets, but proceeds with a single fit.
  • domain assumption The image processing features (grain density, RMS roughness, DoG blob, GLCM dissimilarity) are valid proxies for the physical properties of MoS2 films.
    Section 2.3 computes these with Otsu, Canny, DoG and GLCM using parameters inherited from prior work (Ref 50) and treats the correlations as validation of the latent space.
  • domain assumption The cleaned dataset (only 1 um x 1 um AFM images and only 4 mW Raman/PL spectra, about 30% of the original data) is representative of the synthesis space.
    Section 2.2 applies these filters for consistency, but the effect of the 70% exclusion on generality is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Modal Characterization of Thin Film MoS$_2$ Using Generative Models." pith.science (2026). https://pith.science/paper/FRDPL4DB

@misc{pith2026250524065,
  author       = {Pith},
  title        = {Pith review of: Cross-Modal Characterization of Thin Film MoS$_2$ Using Generative Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FRDPL4DB}},
  note         = {Machine review of arXiv:2505.24065}
}
abstract

The growth and characterization of materials using empirical optimization typically requires a significant amount of expert time, experience, and resources. Several complementary characterization methods are routinely performed to determine the quality and properties of a grown sample. Machine learning (ML) can support the conventional approaches by using historical data to guide and provide speed and efficiency to the growth and characterization of materials. Specifically, ML can provide quantitative information from characterization data that is typically obtained from a different modality. In this study, we have investigated the feasibility of projecting the quantitative metric from microscopy measurements, such as atomic force microscopy (AFM), using data obtained from spectroscopy measurements, like Raman spectroscopy. Generative models were also trained to generate the full and specific features of the Raman and photoluminescence spectra from each other and the AFM images of the thin film MoS$_2$. The results are promising and have provided a foundational guide for the use of ML for the cross-modal characterization of materials for their accelerated, efficient, and cost-effective discovery.

Figures

Figures reproduced from arXiv: 2505.24065 by the authors.

Figure 1
Figure 1. Examples of samples showing the AFM (first row), Raman (second row), and PL [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. The workflow of the study. (1) is the unsupervised representation of the AFM images to obtain latent features. In obtaining latent features, pretrained ResNet152 is used to extract AFM features and projected to fewer dimensions using Uniform Manifold Ap￾proximation and Projection (UMAP) and principal component analysis (PCA). (2a) and (2b) are the regression models of the AFM latent features using the Raman and phot… view at source ↗
Figure 3
Figure 3. Sample Atomic Force Microscopy (AFM) images in the Uniform Manifold Ap [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Predicting the AFM embedding from the Raman spectra. (a) and (b) are the [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Sample held-out Raman spectral generated from the corresponding AFM images [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Sample held-out photoluminescence (PL) spectra generated from the corresponding [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: A1g-E1 2g split obtained from the Raman generated from (a) AFM images and (b) PL spectra. Beyond generating full spectra, we aimed to determine the quantitative features of the generated spectra and compare them with those obtained from true spectra. The important feat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 57 canonical work pages

  1. [1]

    Synthesis and characterization of vertically standing MoS2 nanosheets

    Li, H.; Wu, H.; Yuan, S.; Qian, H. Synthesis and characterization of vertically standing MoS2 nanosheets. Scientific reports 2016, 6, 21171

    work page 2016

  2. [2]

    P.; Keating, C

    Mulvaney, S. P.; Keating, C. D. Raman spectroscopy. Analytical Chemistry 2000, 72, 145--158

    work page 2000

  3. [3]

    A comprehensive review on Raman spectroscopy applications

    Orlando, A.; Franceschini, F.; Muscas, C.; Pidkova, S.; Bartoli, M.; Rovere, M.; Tagliaferro, A. A comprehensive review on Raman spectroscopy applications. Chemosensors 2021, 9, 262

    work page 2021

  4. [4]

    Atomic force microscopy

    Rugar, D.; Hansma, P. Atomic force microscopy. Physics today 1990, 43, 23--30

    work page 1990

  5. [5]

    Giessibl, F. J. Advances in atomic force microscopy. Reviews of modern physics 2003, 75, 949

    work page 2003

  6. [6]

    Atomic force microscopy for two-dimensional materials: A tutorial review

    Zhang, H.; Huang, J.; Wang, Y.; Liu, R.; Huai, X.; Jiang, J.; Anfuso, C. Atomic force microscopy for two-dimensional materials: A tutorial review. Optics Communications 2018, 406, 3--17

    work page 2018

  7. [7]

    W.; Yang, S

    Hong, W.; Park, C.; Shim, G. W.; Yang, S. Y.; Choi, S.-Y. Wafer-Scale Uniform Growth of an Atomically Thin MoS2 Film with Controlled Layer Numbers by Metal--Organic Chemical Vapor Deposition. 2021,

    work page 2021

  8. [8]

    Chemical Reviews 2024, 124, 10112--10191

    Wu, R.; Zhang, H.; Ma, H.; Zhao, B.; Li, W.; Chen, Y.; Liu, J.; Liang, J.; Qin, Q.; Qi, W.; others Synthesis, modulation, and application of two-dimensional TMD heterostructures. Chemical Reviews 2024, 124, 10112--10191

    work page 2024

Show all 59 references
  1. [9]

    Advancements in 2D transition metal dichalcogenides (TMDs) inks for printed optoelectronics: A comprehensive review

    Shahbaz, I.; Tahir, M.; Li, L.; Song, Y. Advancements in 2D transition metal dichalcogenides (TMDs) inks for printed optoelectronics: A comprehensive review. Materials Today 2024,

    2024

  2. [10]

    Z.; Dahshan, A.; Aftab, S.; Hegazy, H

    Ahmed, A.; Iqbal, M. Z.; Dahshan, A.; Aftab, S.; Hegazy, H. H.; others Recent advances in 2D transition metal dichalcogenide-based photodetectors: a review. Nanoscale 2024,

    2024

  3. [11]

    Advances in MoS2-Based Biosensors: From Material Fabrication and Characterization to Biomedical, Environmental, and Industrial Applications

    Lai, C.-L.; Mukundan, A.; Karmakar, R.; Kaur, R.; Huang, K.-L.; Wang, H.-C. Advances in MoS2-Based Biosensors: From Material Fabrication and Characterization to Biomedical, Environmental, and Industrial Applications. Biosensors 2025, 15, 371

    2025

  4. [12]

    Science China Information Sciences 2025, 68, 112401

    Guo, Y.; Wang, Y.; Deng, W.; Wu, Y.; Li, J.; Li, K.; Zhao, Y.; Yu, S.; Wang, X.; Zhang, Y.; others MoS2 synaptic transistor with one-step manufacture. Science China Information Sciences 2025, 68, 112401

    2025

  5. [13]

    MoS2-based quantum dot artificial synapses for neuromorphic computing

    Liu, G.; Liu, H.; Fan, F.; Gu, Y.; Wei, L.; Xiang, X.; Wang, Y.; Li, Q. MoS2-based quantum dot artificial synapses for neuromorphic computing. Materials Today Physics 2025, 53, 101703

    2025

  6. [14]

    Chen, C.; Trainor, N.; Kumari, S.; Myja, H.; K \"u mmell, T.; Zhang, Z.; Zhang, Y.; Bisht, A.; Sadaf, M. U. K.; Sakib, N. U.; others Effect of growth temperature on the microstructure and properties of epitaxial MoS2 monolayers grown by metalorganic chemical vapor deposition. ...

    2024

  7. [15]

    Li, H.; Zhang, Q.; Yap, C. C. R.; Tay, B. K.; Edwin, T. H. T.; Olivier, A.; Baillargeat, D. From bulk to monolayer MoS2: evolution of Raman scattering. Advanced Functional Materials 2012, 22, 1385--1390

    2012

  8. [16]

    Raman spectroscopy characterization of two-dimensional materials

    Liang, F.; Xu, H.; Wu, X.; Wang, C.; Luo, C.; Zhang, J. Raman spectroscopy characterization of two-dimensional materials. Chinese Physics B 2018, 27, 037802

    2018

  9. [17]

    R.; Bertolazzi, S.; Brivio, J.; Watson, M.; Wu, X.; Kis, A.; Luo, T.; Hight Walker, A

    Yan, R.; Simpson, J. R.; Bertolazzi, S.; Brivio, J.; Watson, M.; Wu, X.; Kis, A.; Luo, T.; Hight Walker, A. R.; Xing, H. G. Thermal conductivity of monolayer molybdenum disulfide obtained from temperature-dependent Raman spectroscopy. ACS nano 2014, 8, 986--993

    2014

  10. [18]

    E.; Heinz, T

    Lee, C.; Yan, H.; Brus, L. E.; Heinz, T. F.; Hone, J.; Ryu, S. Anomalous lattice vibrations of single-and few-layer MoS2. ACS nano 2010, 4, 2695--2700

    2010

  11. [19]

    Catal \'a n-G \'o mez, S.; Garg, S.; Redondo-Cubero, A.; Gordillo, N.; de Andr \'e s, A.; Nucciarelli, F.; Kim, S.; Kung, P.; Pau, J. L. Photoluminescence enhancement of monolayer MoS 2 using plasmonic gallium nanoparticles. Nanoscale Advances 2019, 1, 884--893

    2019

  12. [20]

    I.; Golovynska, I.; Li, B.; Lin, D.; Qu, J

    Golovynskyi, S.; Irfan, I.; Bosi, M.; Seravalli, L.; Datsenko, O. I.; Golovynska, I.; Li, B.; Lin, D.; Qu, J. Exciton and trion in few-layer MoS2: Thickness-and temperature-dependent photoluminescence. Applied Surface Science 2020, 515, 146033

    2020

  13. [21]

    Shape-uniform, high-quality monolayered MoS2 crystals for gate-tunable photoluminescence

    Zhang, X.; Nan, H.; Xiao, S.; Wan, X.; Ni, Z.; Gu, X.; Ostrikov, K. Shape-uniform, high-quality monolayered MoS2 crystals for gate-tunable photoluminescence. ACS Applied Materials & Interfaces 2017, 9, 42121--42130

    2017

  14. [22]

    M.; Hossain, L.; Lin, Y.-C.; Piasecki, A

    Eichfeld, S. M.; Hossain, L.; Lin, Y.-C.; Piasecki, A. F.; Kupp, B.; Birdwell, A. G.; Burke, R. A.; Lu, N.; Peng, X.; Li, J.; others Highly scalable, atomically thin WSe2 grown via metal--organic chemical vapor deposition. ACS nano 2015, 9, 2080--2087

    2015

  15. [23]

    Chemical Vapor Deposition of a Single-Crystalline MoS2 Monolayer through Anisotropic 2D Crystal Growth on Stepped Sapphire Surface

    Kandybka, I.; Groven, B.; Medina Silva, H.; Sergeant, S.; Nalin Mehta, A.; Koylan, S.; Shi, Y.; Banerjee, S.; Morin, P.; Delabie, A. Chemical Vapor Deposition of a Single-Crystalline MoS2 Monolayer through Anisotropic 2D Crystal Growth on Stepped Sapphire Surface. ACS nano 202...

    2024

  16. [24]

    A.; Wu, C.; Reinhart, W

    Moses, I. A.; Wu, C.; Reinhart, W. F. Crystal growth characterization of WSe2 thin film using machine learning. Materials Today Advances 2024, 22, 100483

    2024

  17. [25]

    M.; Yesudoss, D.; Johnson, D.; Djire, A

    Yoo, R. M.; Yesudoss, D.; Johnson, D.; Djire, A. A review on the application of in-situ Raman spectroelectrochemistry to understand the mechanisms of hydrogen evolution reaction. ACS Catalysis 2023, 13, 10570--10601

    2023

  18. [26]

    High-temperature in situ investigation of chemical vapor deposition to reveal growth mechanisms of monolayer molybdenum disulfide

    Xue, H.; Wu, G.; Zhao, B.; Wang, D.; Wu, X.; Hu, Z. High-temperature in situ investigation of chemical vapor deposition to reveal growth mechanisms of monolayer molybdenum disulfide. ACS Applied Electronic Materials 2020, 2, 1925--1933

    2020

  19. [27]

    A deep learning approaches in text-to-speech system: a systematic review and recent research perspective

    Kumar, Y.; Koul, A.; Singh, C. A deep learning approaches in text-to-speech system: a systematic review and recent research perspective. Multimedia Tools and Applications 2023, 82, 15171--15197

    2023

  20. [28]

    S.; Rotskoff, G

    Hu, F.; Chen, M. S.; Rotskoff, G. M.; Kanan, M. W.; Markland, T. E. Accurate and efficient structure elucidation from routine one-dimensional NMR spectra using multitask machine learning. ACS Central Science 2024, 10, 2162--2170

    2024

  21. [29]

    Transformer-Based Models for Predicting Molecular Structures from Infrared Spectra Using Patch-Based Self-Attention

    Wu, W.; Leonardis, A.; Jiao, J.; Jiang, J.; Chen, L. Transformer-Based Models for Predicting Molecular Structures from Infrared Spectra Using Patch-Based Self-Attention. The Journal of Physical Chemistry A 2025,

    2025

  22. [30]

    Alberts, M.; Laino, T.; Vaucher, A. C. Leveraging infrared spectroscopy for automated structure elucidation. Communications Chemistry 2024, 7, 268

    2024

  23. [31]

    E.; Chenthamarakshan, V.; Das, P.; Kavraki, L

    Litsa, E. E.; Chenthamarakshan, V.; Das, P.; Kavraki, L. E. An end-to-end deep learning framework for translating mass spectra to de-novo molecules. Communications Chemistry 2023, 6, 132

    2023

  24. [32]

    u hrkop, K.; B \

    Stravs, M. A.; D \"u hrkop, K.; B \"o cker, S.; Zamboni, N. MSNovelist: de novo structure generation from mass spectra. Nature Methods 2022, 19, 865--870

    2022

  25. [33]

    Cross-Modal Prediction of Spectral and Structural Descriptors via a Pretrained Model Enhanced with Chemical Insights

    Yang, G.; Jiang, S.; Luo, Y.; Wang, S.; Jiang, J. Cross-Modal Prediction of Spectral and Structural Descriptors via a Pretrained Model Enhanced with Chemical Insights. The Journal of Physical Chemistry Letters 2024, 15, 8766--8772

    2024

  26. [34]

    J.; Lee, J

    Jeong, J.; Kim, T.; Lee, B. J.; Lee, J. Predicting AFM Topography from Optical Microscope Images Using Deep-Learning. Advanced Intelligent Systems 2023, 5, 2200317

    2023

  27. [35]

    Pair-variational autoencoders for linking and cross-reconstruction of characterization data from complementary structural characterization techniques

    Lu, S.; Jayaraman, A. Pair-variational autoencoders for linking and cross-reconstruction of characterization data from complementary structural characterization techniques. JACS Au 2023, 3, 2510--2521

    2023

  28. [36]

    H.; Chubarov, M.; Xiang, Y.; Jariwala, B.; Zhang, F.; Alem, N.; Wang, G.-C.; Robinson, J

    Zhang, X.; Choudhury, T. H.; Chubarov, M.; Xiang, Y.; Jariwala, B.; Zhang, F.; Alem, N.; Wang, G.-C.; Robinson, J. A.; Redwing, J. M. Diffusion-controlled epitaxy of large area coalesced WSe2 monolayers on sapphire. Nano letters 2018, 18, 1049--1056

    2018

  29. [37]

    A.; Redwing, J.; Curtis, M.; Kim, M.; Trainor, N.; Kumar, S

    Chen, C.; Kowalczyk, D.; Moses, I. A.; Redwing, J.; Curtis, M.; Kim, M.; Trainor, N.; Kumar, S. Data: Cross-Modal Characterization of Thin Film MoS2 Using Generative Models. https://doi.org/10.26207/pbbx-w030

    doi

  30. [38]

    A.; Reinhart, W

    Moses, I. A.; Reinhart, W. Code and Data for Cross-Modal Characterization of Thin Film MoS2 Using Generative Models. 2025; https://doi.org/10.5281/zenodo.15533400

    2025 doi

  31. [39]

    Deep Residual Learning for Image Recognition

    He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. 2015

    2015

  32. [40]

    ImageNet: A large-scale hierarchical image database

    Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; Fei-Fei, L. ImageNet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition. 2009; pp 248--255

    2009

  33. [41]

    Umap: Uniform manifold approximation and projection for dimension reduction

    McInnes, L.; Healy, J.; Melville, J. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 2018,

    2018 arXiv

  34. [42]

    A.; Reinhart, W

    Yu, M.; Moses, I. A.; Reinhart, W. F.; Law, S. Multimodal Machine Learning Analysis of GaSe Molecular Beam Epitaxy Growth Conditions. ACS Applied Materials & Interfaces 2025, 17, 34707--34716, PMID: 40434265

    2025

  35. [43]

    W.; Ng, L

    Becht, E.; McInnes, L.; Healy, J.; Dutertre, C.-A.; Kwok, I. W.; Ng, L. G.; Ginhoux, F.; Newell, E. W. Dimensionality reduction for visualizing single-cell data using UMAP. Nature biotechnology 2019, 37, 38--44

    2019

  36. [44]

    F.R.S., K. P. LIII. On lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 1901, 2, 559--572

    1901

  37. [45]

    T.; Cadima, J

    Jolliffe, I. T.; Cadima, J. Principal component analysis: a review and recent developments. Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences 2016, 374, 20150202

    2016

  38. [46]

    Lowe, D. G. Distinctive image features from scale-invariant keypoints. International journal of computer vision 2004, 60, 91--110

    2004

  39. [47]

    M.; Shanmugam, K.; Dinstein, I

    Haralick, R. M.; Shanmugam, K.; Dinstein, I. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics 1973, SMC-3, 610--621

    1973

  40. [48]

    A Threshold Selection Method from Gray-Level Histograms

    Otsu, N. A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems, Man, and Cybernetics 1979, 9, 62--66

    1979

  41. [49]

    A Computational Approach to Edge Detection

    Canny, J. A Computational Approach to Edge Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 1986, PAMI-8, 679--698

    1986

  42. [50]

    A.; Reinhart, W

    Moses, I. A.; Reinhart, W. F. Transfer learning for multi-material classification of transition metal dichalcogenides with atomic force microscopy. Machine Learning: Science and Technology 2025, 5, 045081

    2025

  43. [51]

    Advances in neural information processing systems 2019, 32

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; others Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 2019, 32

    2019

  44. [52]

    the Journal of machine Learning research 2011, 12, 2825--2830

    Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; others Scikit-learn: Machine learning in Python. the Journal of machine Learning research 2011, 12, 2825--2830

    2011

  45. [53]

    Optuna: A Next-generation Hyperparameter Optimization Framework

    Akiba, T.; Sano, S.; Yanase, T.; Ohta, T.; Koyama, M. Optuna: A Next-generation Hyperparameter Optimization Framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2019

    2019

  46. [54]

    L.; Kung, Y.-C.; Krasnozhon, D.; Chen, M.-W.; others Large-area epitaxial monolayer MoS2

    Dumcenco, D.; Ovchinnikov, D.; Marinov, K.; Lazic, P.; Gibertini, M.; Marzari, N.; Sanchez, O. L.; Kung, Y.-C.; Krasnozhon, D.; Chen, M.-W.; others Large-area epitaxial monolayer MoS2. ACS nano 2015, 9, 4611--4620

    2015

  47. [55]

    Atomically thin MoS2: a versatile nongraphene 2D material

    Venkata Subbaiah, Y.; Saji, K.; Tiwari, A. Atomically thin MoS2: a versatile nongraphene 2D material. Advanced Functional Materials 2016, 26, 2046--2069

    2016

  48. [56]

    W.; N rskov, J

    Lauritsen, J.; Bollinger, M.; L gsgaard, E.; Jacobsen, K. W.; N rskov, J. K.; Clausen, B.; Tops e, H.; Besenbacher, F. Atomic-scale insight into structure and morphology changes of MoS2 nanoclusters in hydrotreating catalysts. Journal of Catalysis 2004, 221, 510--522

    2004

  49. [57]

    Analyzing Fundamental Properties of Two-Dimensional Materials by Raman Spectroscopy from Microscale to Nanoscale

    Wu, H.; Lin, M.-L.; Zhou, Y.; Zhang, X.; Tan, P.-H. Analyzing Fundamental Properties of Two-Dimensional Materials by Raman Spectroscopy from Microscale to Nanoscale. Analytical Chemistry 2023, 95, 10821--10838

    2023

  50. [58]

    Analysis of photoluminescence behavior of high-quality single-layer MoS 2

    Xu, L.; Zhao, L.; Wang, Y.; Zou, M.; Zhang, Q.; Cao, A. Analysis of photoluminescence behavior of high-quality single-layer MoS 2. Nano Research 2019, 12, 1619--1624

    2019

  51. [59]

    u hrkop, Kai and B \

    Moses, I. A.; Reinhart, W. F. Quantitative analysis of MoS2 thin film micrographs with machine learning. Materials Characterization 2024, 209, 113701 mcitethebibliography main.bib0000664000000000000000000006251115050700620011160 0ustar rootroot@article moses2024crystal, title ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.