Identification and Structural Characterization of Twisted Atomically Thin Bilayer Materials by Deep Learning

Eduardo R Hern\'andez; Haitao Yang; Haosen Chen; Heng Wu; Kexin He; Mingming Gong; Ping-Heng Tan; Ruiqi Hu; Wenshuai Hu; Xiaolong He

arxiv: 2604.15960 · v1 · submitted 2026-04-17 · ❄️ cond-mat.mtrl-sci

Identification and Structural Characterization of Twisted Atomically Thin Bilayer Materials by Deep Learning

Haitao Yang , Ruiqi Hu , Heng Wu , Xiaolong He , Yan Zhou , Yizhe Xue , Kexin He , Wenshuai Hu

show 6 more authors

Haosen Chen Mingming Gong Xin Zhang Ping-Heng Tan Eduardo R Hern\'andez Yong Xie

This is my paper

Pith reviewed 2026-05-10 08:48 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords twisted bilayerMoS2deep learningconvolutional neural networktwist anglechemical vapor depositionoptical microscopysemantic segmentation

0 comments

The pith

Convolutional neural networks trained on synthetic images accurately predict twist angles and thicknesses in CVD MoS2 bilayers from optical images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops two convolutional neural network models to analyze optical microscope images of molybdenum disulfide flakes grown by chemical vapor deposition. One model performs semantic segmentation to determine flake thicknesses, while the second predicts the twist angle in bilayer regions. By training the twist angle model on over 10,000 synthetic images covering various shapes from hexagonal to triangular, the approach enables rapid, automated characterization that is validated against second harmonic generation and Raman spectroscopy measurements. This matters because manual inspection of twisted 2D materials is slow and subjective, limiting the ability to scale up studies of their unique properties for electronics applications.

Core claim

A semantic segmentation CNN identifies MoS2 flake thicknesses from color space in optical microscopy, and a second CNN, trained on 10,000 synthetic images of hexagonal to triangular bilayers, predicts twist angles with accuracy confirmed by SHG and Raman spectroscopy, enabling scalable automated inspection of twisted atomically thin CVD-grown bilayers.

What carries the argument

Two CNN models—one for semantic segmentation of thicknesses and one for twist angle prediction from optical images—trained on a large synthetic dataset of bilayer geometries.

If this is right

High-throughput screening of large numbers of twisted bilayer samples becomes feasible without constant spectroscopic checks.
Initial mapping of twist-angle distributions across CVD-grown wafers can be automated for process optimization.
Layer thickness identification can be integrated into standard optical inspection workflows for 2D material quality control.
The method supports rapid selection of specific twist angles for devices that exploit twist-dependent electronic or optical effects.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same synthetic-image training strategy could be adapted to other layered materials where optical contrast varies with stacking.
Combining the CNN outputs with automated stage movement on a microscope would allow wafer-scale mapping in hours rather than days.
If the models generalize, they could flag candidate regions for further study with more precise but slower techniques like ARPES.

Load-bearing premise

The 10,000 synthetic images spanning hexagonal to triangular geometries sufficiently represent the optical appearance and variability of real CVD-grown bilayer MoS2 flakes under the microscope.

What would settle it

A new collection of real bilayer MoS2 flakes whose twist angles are measured independently by second harmonic generation spectroscopy, where the CNN predictions show systematic disagreement beyond the initial validation set.

Figures

Figures reproduced from arXiv: 2604.15960 by Eduardo R Hern\'andez, Haitao Yang, Haosen Chen, Heng Wu, Kexin He, Mingming Gong, Ping-Heng Tan, Ruiqi Hu, Wenshuai Hu, Xiaolong He, Xin Zhang, Yan Zhou, Yizhe Xue, Yong Xie.

**Figure 2.** Figure 2: Segmentation techniques for classifying thickness in atomically thin CVD-grown [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Deep learning approach for recognizing twist angles in atomically thin bilayer [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Performance evaluation of the twisted bilayer MoS [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 4.** Figure 4: The identification of twist bilayer MoS2 using OpenCV is also demonstrated in the supplementary (Figure S8) and shown in the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: Moir´e phonons in twisted CVD-grown bilayer MoS [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

read the original abstract

Two-dimensional materials are expected to play an important role in next-generation electronics and optoelectronic devices. Recently, twisted bilayer graphene and transition metal dichalcogenides have attracted significant attention due to their unique physical properties and potential applications. In this study we describe the use of optical microscopy to collect the color space of chemical vapor deposition (CVD) molybdenum disulfide ($\mbox{MoS}_2$), and the application of a semantic segmentation convolutional neural network (CNN) to accurately and rapidly identify thicknesses of $\mbox{MoS}_2$ flakes. A second CNN model is trained to provide precise predictions on the twist angle of CVD-grown bilayer flakes. This model harnessed a dataset comprising over 10,000 synthetic images, encompassing geometries spanning from hexagonal to triangular shapes. Subsequent validation of the deep learning predictions on twist angles was executed through the second harmonic generation and Raman spectroscopy. Our results introduce a scalable methodology for automated inspection of twisted atomically thin CVD-grown bilayer.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A practical dual-CNN pipeline for thickness and twist-angle screening in CVD MoS2 bilayers, trained on large synthetic data and checked with SHG and Raman.

read the letter

The paper's main offering is a straightforward automated pipeline: one CNN does semantic segmentation to pick out MoS2 flake thicknesses from optical images, and a second CNN predicts the twist angle after training on more than 10,000 synthetic images that vary flake shapes from hexagonal to triangular. They then test the twist predictions on real CVD samples using second-harmonic generation and Raman spectroscopy. That combination is the concrete step forward here, since earlier work has used CNNs on 2D images but not at this scale for twist angles in bilayer MoS2 specifically. The synthetic-data choice is sensible for getting enough labeled examples without manual annotation of every real flake, and the independent spectroscopic validation keeps the claims from being circular. The method should help labs that grow lots of these bilayers and need faster screening than running full spectroscopy on every piece. It targets experimental groups working on twisted TMDs for electronics or optoelectronics who want quicker optical-based checks. One clear soft spot is the lack of reported accuracy numbers or error bars in the abstract; without those, it is hard to judge whether the twist predictions are reliable enough for routine lab use or if they drift on certain flake types. The assumption that the synthetic set captures the full range of real microscope appearances also needs checking in the full text, since CVD flakes can show extra defects or contrast variations not present in clean simulations. Minor gaps like that do not sink the work, but they mean the paper would benefit from more quantitative tables and perhaps tests on additional materials. This is worth sending for peer review. The core idea is grounded, the validation is independent, and the field can use better automation tools even if the current version needs tighter performance metrics before widespread adoption.

Referee Report

2 major / 3 minor

Summary. The manuscript describes a deep learning pipeline for automated characterization of CVD-grown MoS2 flakes: a semantic segmentation CNN identifies flake thicknesses from optical microscopy images, while a second CNN predicts twist angles in bilayer regions. The twist-angle model is trained on >10,000 synthetic images spanning hexagonal-to-triangular geometries; predictions are validated on real samples using second-harmonic generation and Raman spectroscopy. The central claim is that this approach provides a scalable, automated inspection method for twisted atomically thin bilayers.

Significance. If the reported accuracy and generalization hold, the work offers a practical, high-throughput alternative to labor-intensive manual or spectroscopic characterization of twist angles in 2D materials. The combination of synthetic-data training with independent experimental validation (SHG/Raman) is a methodological strength that could accelerate research on twistronics and CVD-grown heterostructures.

major comments (2)

[Methods, Dataset Generation] Methods, Dataset Generation section: The assertion that the >10,000 synthetic images 'sufficiently represent' the optical appearance and variability of real CVD-grown bilayer MoS2 flakes is load-bearing for the generalization claim, yet no quantitative metrics (e.g., distribution overlap in color histograms, edge sharpness, or contrast statistics) comparing synthetic versus experimental images are provided. This leaves open whether domain shift affects twist-angle prediction accuracy on real samples.
[Results, Validation] Results, Validation subsection: While SHG and Raman are cited as independent validation, the manuscript does not report quantitative agreement metrics (e.g., mean absolute error or correlation coefficient between CNN-predicted and spectroscopically measured twist angles) or the number of flakes tested. Without these, it is difficult to assess whether the model meets the 'precise predictions' standard claimed in the abstract.

minor comments (3)

[Abstract] Abstract: The phrase 'collect the color space' is imprecise; specify the color space (RGB, HSV, LAB) and any preprocessing steps applied to the optical images.
[Figures] Figure captions: Several figures lack scale bars or explicit indication of which images are synthetic versus experimental, reducing clarity for readers attempting to reproduce the workflow.
[Introduction] Notation: The manuscript uses 'twist angle' without consistently defining the reference (e.g., relative to armchair or zigzag direction); a brief clarification in the introduction would aid non-specialist readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and recommendation for minor revision. We address the two major comments below and will revise the manuscript to incorporate the requested quantitative details.

read point-by-point responses

Referee: [Methods, Dataset Generation] Methods, Dataset Generation section: The assertion that the >10,000 synthetic images 'sufficiently represent' the optical appearance and variability of real CVD-grown bilayer MoS2 flakes is load-bearing for the generalization claim, yet no quantitative metrics (e.g., distribution overlap in color histograms, edge sharpness, or contrast statistics) comparing synthetic versus experimental images are provided. This leaves open whether domain shift affects twist-angle prediction accuracy on real samples.

Authors: We agree that quantitative metrics would better support the claim that the synthetic dataset sufficiently represents real CVD-grown flakes. In the revised manuscript we will add direct comparisons, including color histogram overlap, edge sharpness, and contrast statistics between the synthetic images and experimental optical micrographs. We will also report the twist-angle prediction accuracy of the model when evaluated on a set of real bilayer images to quantify any domain shift. revision: yes
Referee: [Results, Validation] Results, Validation subsection: While SHG and Raman are cited as independent validation, the manuscript does not report quantitative agreement metrics (e.g., mean absolute error or correlation coefficient between CNN-predicted and spectroscopically measured twist angles) or the number of flakes tested. Without these, it is difficult to assess whether the model meets the 'precise predictions' standard claimed in the abstract.

Authors: We acknowledge that explicit quantitative metrics and the sample size are needed to substantiate the validation claims. The revised manuscript will state the number of bilayer flakes characterized by SHG and Raman spectroscopy and will report quantitative agreement statistics, specifically the mean absolute error and the correlation coefficient between the CNN-predicted twist angles and the spectroscopically measured values. revision: yes

Circularity Check

0 steps flagged

No circularity in the ML training and validation pipeline

full rationale

The paper trains semantic segmentation and twist-angle CNNs on a dataset of >10,000 synthetic images and validates predictions on real CVD MoS2 flakes using independent second-harmonic generation and Raman spectroscopy. No equations, fitted parameters, or self-citations are described that would reduce any reported prediction to the training inputs by construction. The central claim (scalable automated inspection) rests on data-driven learning plus external spectroscopic ground truth rather than any self-referential loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard assumptions of convolutional neural networks (local connectivity, translation invariance) and the premise that optical color space plus synthetic geometry suffice to capture real flake appearance; no new physical axioms or entities are introduced.

axioms (1)

domain assumption Convolutional neural networks can learn to map optical color features to physical thickness and twist angle when trained on appropriate data.
Invoked implicitly by training the two CNN models on real and synthetic images.

pith-pipeline@v0.9.0 · 5514 in / 1286 out tokens · 24796 ms · 2026-05-10T08:48:25.152430+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

Unconventional superconductivity in magic-angle graphene superlattices

(1) Cao, Y.; Fatemi, V.; Fang, S.; Watanabe, K.; Taniguchi, T.; Kaxiras, E.; Jarillo- Herrero, P. Unconventional superconductivity in magic-angle graphene superlattices. Nature2018,556, 43–50. (2) Cao, Y.; Fatemi, V.; Demir, A.; Fang, S.; Tomarken, S. L.; Luo, J. Y.; Sanchez- Yamagishi, J. D.; Watanabe, K.; Taniguchi, T.; Kaxiras, E.; Ashoori, R. C.; Jari...

work page 2018
[2]

N.; Bockrath, M

(7) Lau, C. N.; Bockrath, M. W.; Mak, K. F.; Zhang, F. Reproducibility in the fabrication and physics of moir´ e materials.Nature2022,602, 41–50. (8) Quan, J. et al. Phonon renormalization in reconstructed MoS 2 moir´ e superlattices.Nat. Mater.2021,20, 1100–1105. (9) Marcellina, E.; Liu, X.; Hu, Z.; Fieramosca, A.; Huang, Y.; Du, W.; Liu, S.; Zhao, J.; W...

work page 2021
[3]

F.; Thygesen, K

(11) Peimyoo, N.; Deilmann, T.; Withers, F.; Escolar, J.; Nutting, D.; Taniguchi, T.; Watan- abe, K.; Taghizadeh, A.; Craciun, M. F.; Thygesen, K. S.; Russo, S. Electrical tuning of optically active interlayer excitons in bilayer MoS 2.Nat. Nanotechnol.2021,16, 888–893. 17 (12) Sun, L.; Wang, Z.; Wang, Y.; Zhao, L.; Li, Y.; Chen, B.; Huang, S.; Zhang, S.;...

work page 2021
[4]

Controllable growth of monolayer MoS 2 by chemical vapor deposition via close MoO2 precursor for electrical and optical applications.Nanotechnology2017, 28, 084001

(13) Xie, Y.; Wang, Z.; Zhan, Y.; Zhang, P.; Wu, R.; Jiang, T.; Wu, S.; Wang, H.; Zhao, Y.; Nan, T.; Ma, X. Controllable growth of monolayer MoS 2 by chemical vapor deposition via close MoO2 precursor for electrical and optical applications.Nanotechnology2017, 28, 084001. (14) Wang, Z.; Xie, Y.; Wang, H.; Wu, R.; Nan, T.; Zhan, Y.; Sun, J.; Jiang, T.; Zha...

work page 2020
[5]

Transition metal dichalcogenides bilayer single crystals by reverse-flow chemical vapor epitaxy

(16) Zhang, X.; Nan, H.; Xiao, S.; Wan, X.; Gu, X.; Du, A.; Ni, Z.; Ostrikov, K. Transition metal dichalcogenides bilayer single crystals by reverse-flow chemical vapor epitaxy. Nat. Commun.2019,10,

work page 2019
[6]

2H/1T ′ phase WS 2(1−x)Te 2x alloys grown by chemical vapor deposition with tunable band structures.Applied Surface Science2020,504, 144371

(17) Wang, Z.; Sun, J.; Wang, H.; Lei, Y.; Xie, Y.; Wang, G.; Zhao, Y.; Li, X.; Xu, H.; Yang, X.; Feng, L.; Ma, X. 2H/1T ′ phase WS 2(1−x)Te 2x alloys grown by chemical vapor deposition with tunable band structures.Applied Surface Science2020,504, 144371. (18) Dumcenco, D.; Ovchinnikov, D.; Marinov, K.; Lazic, P.; Gibertini, M.; Marzari, N.; Sanchez, O. L...

work page 2014
[7]

Intelligent identification of two-dimensional nanostructures by machine-learning optical microscopy.Nano Res.2018,11, 6316–6324

(36) Lin, X.; Si, Z.; Fu, W.; Yang, J.; Guo, S.; Cao, Y.; Zhang, J.; Wang, X.; Liu, P.; Jiang, K.; Zhao, W. Intelligent identification of two-dimensional nanostructures by machine-learning optical microscopy.Nano Res.2018,11, 6316–6324. (37) Masubuchi, S.; Machida, T. Classifying optical microscope images of exfoliated graphene flakes by data-driven machi...

work page 2018
[8]

I.-J.; Ma, Q.; Cao, Y.; Rodan-Legrain, D.; et al

(38) Han, B.; Lin, Y.; Yang, Y.; Mao, N.; Li, W.; Wang, H.; Yasuda, K.; Wang, X.; Fatemi, V.; Zhou, L.; Wang, J. I.-J.; Ma, Q.; Cao, Y.; Rodan-Legrain, D.; et al. Deep-learning-enabled fast optical identification and characterization of 2D materials. Adv. Mater.2020,32, 2000953. (39) Masubuchi, S.; Watanabe, E.; Seo, Y.; Okazaki, S.; Sasagawa, T.; Watanab...

work page 2020
[9]

M.; Haley, K

(40) Sterbentz, R. M.; Haley, K. L.; Island, J. O. Universal image segmentation for optical identification of 2D materials.Sci. Rep.2021,11,

work page 2021
[10]

Rethinking atrous convolution for semantic image segmentation

(42) Chen, L.-C.; Papandreou, G.; Schroff, F.; Adam, H. Rethinking atrous convolution for semantic image segmentation. 2017; accessed on December 06,

work page 2017
[11]

Fully convolutional networks for semantic segmen- tation

21 (43) Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmen- tation. Proceedings of the IEEE conference on computer vision and pattern recognition. 2015; pp 3431–3440. (44) Howard, A.; Sandler, M.; Chu, G.; Chen, L.-C.; Chen, B.; Tan, M.; Wang, W.; Zhu, Y.; Pang, R.; Vasudevan, V.; Le, Q. V.; Adam, H. Searching for mobile...

work page 2015
[12]

(46) TorchVision maintainers and contributors TorchVision: PyTorch’s Computer Vision Library.https://github.com/pytorch/vision, 2016; Accessed: February 10,

2015; pp 234–241. (46) TorchVision maintainers and contributors TorchVision: PyTorch’s Computer Vision Library.https://github.com/pytorch/vision, 2016; Accessed: February 10,

work page 2015
[13]

Deep residual learning for image recognition

(47) He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. Pro- ceedings of the IEEE conference on computer vision and pattern recognition. 2016; pp 770–778. (48) Lau, C. N.; Bockrath, M. W.; Mak, K. F.; Zhang, F. Reproducibility in the fabrication and physics of moir´ e materials.Nature2022,602, 41–50. (49) Lin, K.-Q.; Holler,...

work page 2016

[1] [1]

Unconventional superconductivity in magic-angle graphene superlattices

(1) Cao, Y.; Fatemi, V.; Fang, S.; Watanabe, K.; Taniguchi, T.; Kaxiras, E.; Jarillo- Herrero, P. Unconventional superconductivity in magic-angle graphene superlattices. Nature2018,556, 43–50. (2) Cao, Y.; Fatemi, V.; Demir, A.; Fang, S.; Tomarken, S. L.; Luo, J. Y.; Sanchez- Yamagishi, J. D.; Watanabe, K.; Taniguchi, T.; Kaxiras, E.; Ashoori, R. C.; Jari...

work page 2018

[2] [2]

N.; Bockrath, M

(7) Lau, C. N.; Bockrath, M. W.; Mak, K. F.; Zhang, F. Reproducibility in the fabrication and physics of moir´ e materials.Nature2022,602, 41–50. (8) Quan, J. et al. Phonon renormalization in reconstructed MoS 2 moir´ e superlattices.Nat. Mater.2021,20, 1100–1105. (9) Marcellina, E.; Liu, X.; Hu, Z.; Fieramosca, A.; Huang, Y.; Du, W.; Liu, S.; Zhao, J.; W...

work page 2021

[3] [3]

F.; Thygesen, K

(11) Peimyoo, N.; Deilmann, T.; Withers, F.; Escolar, J.; Nutting, D.; Taniguchi, T.; Watan- abe, K.; Taghizadeh, A.; Craciun, M. F.; Thygesen, K. S.; Russo, S. Electrical tuning of optically active interlayer excitons in bilayer MoS 2.Nat. Nanotechnol.2021,16, 888–893. 17 (12) Sun, L.; Wang, Z.; Wang, Y.; Zhao, L.; Li, Y.; Chen, B.; Huang, S.; Zhang, S.;...

work page 2021

[4] [4]

Controllable growth of monolayer MoS 2 by chemical vapor deposition via close MoO2 precursor for electrical and optical applications.Nanotechnology2017, 28, 084001

(13) Xie, Y.; Wang, Z.; Zhan, Y.; Zhang, P.; Wu, R.; Jiang, T.; Wu, S.; Wang, H.; Zhao, Y.; Nan, T.; Ma, X. Controllable growth of monolayer MoS 2 by chemical vapor deposition via close MoO2 precursor for electrical and optical applications.Nanotechnology2017, 28, 084001. (14) Wang, Z.; Xie, Y.; Wang, H.; Wu, R.; Nan, T.; Zhan, Y.; Sun, J.; Jiang, T.; Zha...

work page 2020

[5] [5]

Transition metal dichalcogenides bilayer single crystals by reverse-flow chemical vapor epitaxy

(16) Zhang, X.; Nan, H.; Xiao, S.; Wan, X.; Gu, X.; Du, A.; Ni, Z.; Ostrikov, K. Transition metal dichalcogenides bilayer single crystals by reverse-flow chemical vapor epitaxy. Nat. Commun.2019,10,

work page 2019

[6] [6]

2H/1T ′ phase WS 2(1−x)Te 2x alloys grown by chemical vapor deposition with tunable band structures.Applied Surface Science2020,504, 144371

(17) Wang, Z.; Sun, J.; Wang, H.; Lei, Y.; Xie, Y.; Wang, G.; Zhao, Y.; Li, X.; Xu, H.; Yang, X.; Feng, L.; Ma, X. 2H/1T ′ phase WS 2(1−x)Te 2x alloys grown by chemical vapor deposition with tunable band structures.Applied Surface Science2020,504, 144371. (18) Dumcenco, D.; Ovchinnikov, D.; Marinov, K.; Lazic, P.; Gibertini, M.; Marzari, N.; Sanchez, O. L...

work page 2014

[7] [7]

Intelligent identification of two-dimensional nanostructures by machine-learning optical microscopy.Nano Res.2018,11, 6316–6324

(36) Lin, X.; Si, Z.; Fu, W.; Yang, J.; Guo, S.; Cao, Y.; Zhang, J.; Wang, X.; Liu, P.; Jiang, K.; Zhao, W. Intelligent identification of two-dimensional nanostructures by machine-learning optical microscopy.Nano Res.2018,11, 6316–6324. (37) Masubuchi, S.; Machida, T. Classifying optical microscope images of exfoliated graphene flakes by data-driven machi...

work page 2018

[8] [8]

I.-J.; Ma, Q.; Cao, Y.; Rodan-Legrain, D.; et al

(38) Han, B.; Lin, Y.; Yang, Y.; Mao, N.; Li, W.; Wang, H.; Yasuda, K.; Wang, X.; Fatemi, V.; Zhou, L.; Wang, J. I.-J.; Ma, Q.; Cao, Y.; Rodan-Legrain, D.; et al. Deep-learning-enabled fast optical identification and characterization of 2D materials. Adv. Mater.2020,32, 2000953. (39) Masubuchi, S.; Watanabe, E.; Seo, Y.; Okazaki, S.; Sasagawa, T.; Watanab...

work page 2020

[9] [9]

M.; Haley, K

(40) Sterbentz, R. M.; Haley, K. L.; Island, J. O. Universal image segmentation for optical identification of 2D materials.Sci. Rep.2021,11,

work page 2021

[10] [10]

Rethinking atrous convolution for semantic image segmentation

(42) Chen, L.-C.; Papandreou, G.; Schroff, F.; Adam, H. Rethinking atrous convolution for semantic image segmentation. 2017; accessed on December 06,

work page 2017

[11] [11]

Fully convolutional networks for semantic segmen- tation

21 (43) Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmen- tation. Proceedings of the IEEE conference on computer vision and pattern recognition. 2015; pp 3431–3440. (44) Howard, A.; Sandler, M.; Chu, G.; Chen, L.-C.; Chen, B.; Tan, M.; Wang, W.; Zhu, Y.; Pang, R.; Vasudevan, V.; Le, Q. V.; Adam, H. Searching for mobile...

work page 2015

[12] [12]

(46) TorchVision maintainers and contributors TorchVision: PyTorch’s Computer Vision Library.https://github.com/pytorch/vision, 2016; Accessed: February 10,

2015; pp 234–241. (46) TorchVision maintainers and contributors TorchVision: PyTorch’s Computer Vision Library.https://github.com/pytorch/vision, 2016; Accessed: February 10,

work page 2015

[13] [13]

Deep residual learning for image recognition

(47) He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. Pro- ceedings of the IEEE conference on computer vision and pattern recognition. 2016; pp 770–778. (48) Lau, C. N.; Bockrath, M. W.; Mak, K. F.; Zhang, F. Reproducibility in the fabrication and physics of moir´ e materials.Nature2022,602, 41–50. (49) Lin, K.-Q.; Holler,...

work page 2016