Machine learning assisted speckle and OAM spectrum analysis for enhanced turbulence characterisation

Andrew Forbes; Jiangting Li; Junli Wang; Lixin Guo; Mingjian Cheng; Wenjie Jiang; Xiang Yi

REVIEW 2 major objections 5 minor 30 references

Machine learning assisted speckle and OAM spectrum analysis for enhanced turbulence characterisation

T0 review · 2 major / 5 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 a dual-output convolutional network fusing speckle and OAM spectral data classifies simulated atmospheric turbulence regimes with over 80% validation accuracy, about 13 points above the intensity-only baseline.

desk verdict A credible simulation study showing multimodal fusion of speckle and OAM spectra improves turbulence classification, but the headline accuracy is likely inflated by test-set selection and the method lacks experimental validation. read the letter →

arxiv 2505.21878 v3 pith:ITMRIEGK submitted 2025-05-28 physics.optics

Wenjie Jiang , Mingjian Cheng , Lixin Guo , Xiang Yi , Jiangting Li , Junli Wang , Andrew Forbes This is my paper

classification physics.optics

keywords atmosphericturbulencefree-spaceopticalcommunicationorbitalangularmomentumspecklepatternOAMspectrumdeeplearningmultimodalfusionFriedparameter

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 sets out to show that the distortions a turbulent atmosphere imprints on a vortex beam are readable data, not just noise. It trains convolutional neural networks to classify 20 turbulence conditions, labeled by Fried parameter $r_0$ and Reynolds number $Re$, from two views of the same beam after propagation: the far-field intensity speckle pattern and the orbital angular momentum (OAM) spectrum. The best design, a dual-input, dual-output multi-task network, reaches a mean validation accuracy of 80.68%, compared with 67.41% for an intensity-only baseline, and its run-to-run variance is an order of magnitude smaller. The practical stake is that existing free-space optical links and structured-light sensors could double as turbulence monitors without new hardware.

What carries the argument

The load-bearing object is a dual-input, dual-output convolutional network built on an Inception-style backbone, where parallel branches of convolutions with different kernel sizes capture multi-scale speckle texture. The shared backbone maps a $150\times150$ grayscale speckle image to a 128-dimensional feature vector; this is concatenated with the 40-dimensional OAM spectrum, containing the power in topological charges $-20$ through $+19$ obtained by spiral-spectrum expansion, to form a 168-dimensional fused vector. A primary classification head predicts one of 20 turbulence classes from the fused vector, while an auxiliary head classifies from the image features alone and contributes to the total loss $L_{total}=w_1 L_{main}+w_2 L_{aux}$ with $w_1=0.7$ and $w_2=0.4$. This auxiliary branch is the mechanism that carries the paper's stability claim: it imposes an inductive bias on the shared feature extractor so the fused representation cannot overfit to coincidences between the two modalities.

What would settle it

Feed the trained dual-output model experimentally measured speckle patterns and OAM spectra from a laboratory turbulence chamber whose Fried parameter and Reynolds number are independently characterized. If classification accuracy on these real measurements falls to chance or drops far below the simulated 80.68%, while a model retrained on a mix of simulated and real data recovers, the central transfer claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that spatial and modal information are complementary for turbulence inference, and that a multi-task network which fuses them learns representations that are both more accurate and more stable than either modality alone. In 30 independent training runs on a simulated dataset of 4800 samples spanning $r_0$ from 5 to 20 mm and $Re$ from 1574 to 20000, the dual-output fusion model (Model 3) achieved mean validation accuracy 80.68% and median test accuracy 80.75%, against 67.41% and 68.19% for the intensity-only CNN (Model 1), with validation variance dropping from $4.97 \times 10^{-3}$ to $3.96 \times 10^{-4}$. The auxiliary intensity-only classification head is presented as the key to this stability: by forcing the shared backbone to remain discriminative on the unimodal task, it regularizes the features used by the primary fused branch. The paper also shows that the benefit is not generic: a single-output fusion model is less accurate (78.11%) and, unlike the dual-output model, does not degrade when the OAM spectrum is truncated, indicating that the auxiliary branch makes the network actually rely on the phase-derived modality.

Load-bearing premise

The load-bearing premise is that computer-generated random phase screens reproduce what real air does to a light beam, so accuracy learned in simulation will carry over to real turbulence sensing.

Editorial extensions

If this is right

A single post-propagation measurement pair consisting of one speckle image and one OAM spectrum is enough to identify the turbulence regime among 20 classes, bypassing analytic inverse-scattering solutions.
The auxiliary intensity-only head is doing real work: without it, the fusion model is less accurate and stops depending on the OAM spectrum, so multi-task regularization is a viable alternative to dropout or data augmentation.
The probe beam's topological charge matters for sensing, with $l=3$ giving the best accuracy in the tested range from $l=1$ to $5$.
Because training used only 160 samples per class and 15 epochs, the approach is fitted to data-constrained field deployments.
The method classifies $r_0$ and $Re$ jointly, which is a step toward adaptive free-space optical links that adjust to the channel state they infer from the received beam.

Reading between the lines

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

If the simulation-to-reality transfer holds, the same network could be trained on data from existing free-space optical terminals, converting communication receivers into turbulence sensors with no additional optics.
A controlled ablation the paper does not run is to vary $w_1$ and $w_2$ across the full range and measure test accuracy, which would show whether the auxiliary branch helps by regularization or simply by adding extra gradient signal.
The classification setup could be extended to regression of continuous $r_0$ and $Re$ values, which would make the framework usable for dynamically evolving turbulence rather than regime identification alone.
The decisive test, flagged by the authors as future work, is laboratory or field validation: until the model sees experimentally generated speckle and OAM spectra, the 80% accuracy remains a property of the simulation ensemble.

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.

Desk Editor's Note

read the letter

Real content: combining post-propagation intensity speckle with OAM spectra in a dual-output multi-task CNN is a legitimate extension of prior single-modality ML turbulence sensing. The experiments are internally consistent: 30 runs per model, means and variances, confusion matrices, and an ablation of the OAM spectral range. The qualitative claim that fusion outperforms intensity-only classification is plausible and probably correct.

The main soft spot is hyperparameter selection on the test set. Figures 9 and 10(a) plot test accuracy against incident topological charge and OAM spectral range; the chosen values (l=3 and forty modes) are the ones that look best on the test data. The paper then reports the resulting accuracy as the expected performance, without an independent validation set for these choices. That is a multiple-comparisons leak, and the 30-run statistics are conditional on the selected configuration, so they do not fix the bias. The headline 80.68% should be treated as an upper bound, not an unbiased estimate.

Second, simulation-to-reality transfer is untested. The conclusion explicitly lists real or laboratory experiments as future work, which is honest, but it means the reported accuracy is simulation-specific. Third, no code or data are released, so reproducibility is limited. A minor point: the loss weights were also tuned on validation accuracy (Fig. 6), but that is a small grid and less concerning.

The paper is for researchers working on ML-based turbulence sensing with structured light. It is a modest, incremental advance, not a breakthrough. It deserves peer review because the architecture is sensible and the experiments are decent, but the authors should be required to use a nested validation or a separate development set for configuration choices, and to state the selection explicitly. Adding laboratory data or a clear discussion of the simulation-to-experiment gap would materially strengthen it. My verdict would be major revision, not rejection.

Referee Report

2 major / 5 minor

Summary. The paper proposes a machine-learning framework for atmospheric turbulence characterization in free-space optical channels by fusing post-propagation intensity speckle patterns with OAM spectral data. The authors simulate vortex beam propagation through atmospheric turbulence using the split-step phase-screen method with the Von Kármán spectrum, generating a dataset of 4800 samples spanning 20 turbulence classes defined by the Fried parameter r0 and Reynolds number Re. They compare three CNN architectures: an intensity-only baseline (Model 1), a single-output fusion model (Model 2), and a dual-output multi-task fusion model (Model 3). Based on 30 independent runs per model, they report that Model 3 achieves 80.68% mean validation accuracy, outperforming Model 1 (67.41%) and Model 2 (78.11%), with reduced variance. Additional analyses include confusion matrices, precision/recall/F1, and ablations over incident OAM charge and OAM spectral range.

Significance. If the reported results are unbiased, the paper demonstrates a practical, data-efficient fusion strategy for inferring turbulence parameters from structured-light measurements, with a small per-class sample size and low computational overhead. The study's strengths include 30-run statistics, class-wise confusion matrices, and explicit ablations that indicate the OAM spectrum contributes useful information beyond intensity features. The central qualitative claim—that multimodal fusion aids turbulence classification—is plausible and likely robust to the methodological concerns raised below. However, the headline quantitative result is compromised by an apparent test-set selection issue that needs to be resolved before the specific accuracy figures can be trusted.

major comments (2)

[Section 4, Figs. 9 and 10(b)] The primary experiments adopt an incident OAM topological charge of l=3 and an OAM spectral range of [-20, +19] after these values were selected based on test-set performance curves. Fig. 9 reports test accuracy for l=1..5 and states that all models peak at l=3, and Fig. 10(b) reports test accuracy versus spectral range, with the full range 'adopted for all primary experiments.' Since the same test set is later used for the reported 30-run statistics in Fig. 7(c), the headline 80.68% validation accuracy (Model 3) and the test accuracies in Fig. 7(c) are conditional on hyperparameters chosen using the test set. This is a test-set leakage / multiple-comparisons problem that can inflate the reported performance. Please either perform hyperparameter selection on a separate validation split or use nested cross-validation, and report both the selected and the corresponding unbiased accuracy estimates.
[Section 3, Fig. 6] The loss weights w1=0.7 and w2=0.4 are 'empirically found' to yield the best accuracy based on Fig. 6, which plots accuracy over 30 runs for different weight combinations. The manuscript does not state whether this tuning was performed on the validation split, the test split, or a separate holdout. If the test set was involved in any way, the final accuracy estimate is again subject to selection bias. Please clarify the tuning protocol and, if the test set was used, provide corrected results or a sensitivity analysis over plausible weights.

minor comments (5)

[Section 2, Eq. (2)] Equation (2) appears garbled in typesetting: the expression '13.86 1 1 sin tan 12 66' is missing operators and arguments, making the formula unreadable. Please check the equation against the original source.
[Section 2, OAM spectral range description] The text says 'Measurements covered topological charges from -20 to +19 (40 discrete OAM modes),' but the incident beam is l=3. Clarify that the OAM spectrum is obtained by decomposing the output field over a range of topological charges, not that multiple incident beams were used.
[Section 4, Fig. 10(b) and explanation] The explanation that Model 3 degrades under OAM truncation because its architecture has 'a dedicated branch processes the OAM spectrum' is inaccurate: both Model 2 and Model 3 concatenate OAM data to image features in their classification branches. The difference in sensitivity between Model 2 and Model 3 is not explained by the stated architecture. Please revise the explanation or provide an alternative mechanism.
[Section 2, Reynolds number definition] The relation Re=(L0/l0)^(4/3) is introduced without a citation or derivation. Since the Reynolds number is here used as a classification label and is not a standard atmospheric-optics metric, a reference or justification would help the reader assess the physical interpretation.
[References, [1]] Reference 1 contains an incomplete page range ('PP, 1–10'); please update it to the final pagination.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: reported accuracies are held-out test results on simulated data with physically defined labels; configuration choices (l=3, spectral range) were fixed at dataset construction, and self-citations are non-load-bearing.

full rationale

Walk of the derivation chain: (1) The physical forward model (LG beam with l=3 propagated through Von Kármán phase screens via split-step Fourier method, Table 2: 'OAM mode l = 3') generates intensity speckle and OAM spectra from turbulence parameters (r0, Re); the labels are simulation inputs, not model outputs, so no self-definitional loop exists. (2) The central numerical claim — Model 3 validation accuracy 80.68% vs 67.41% for Model 1 and median test accuracy 80.75% — is evaluated on a held-out test split ('randomly partitioned into training, validation, and test subsets using a 4:1:1 ratio') over 30 independent runs, so it is not a fitted parameter renamed as a prediction. (3) The configuration choices a skeptic might flag (l=3 and OAM range [-20,+19]) were fixed at dataset construction ('Measurements covered topological charges from -20 to +19'), not selected on the test set; the l=1..5 sweep (Fig. 9) and spectral-range ablation (Fig. 10) are post-hoc consistency checks, and the full range was adopted because wider spectra carry more information (Fig. 10a), not to force the headline number. (4) Loss weights (w1=0.7, w2=0.4) are empirically tuned hyperparameters and are not used to fabricate a law or normalize a result into existence. (5) Citations of co-authored work (refs [4], [28], [29], including authors Cheng and Forbes) are used for implementation details ('For implementation details, refer to [28]') and motivation; they are not load-bearing uniqueness premises and do not forbid alternatives. (6) The conclusion explicitly states the main limitation: 'Future research directions should focus on extending this methodology to include validation against real or simulated laboratory-based turbulence experiments', so the simulation-to-reality gap is acknowledged rather than smuggled in. Residual non-circularity concern: reporting 'validation accuracy' in the abstract versus test accuracy in Section 4, and the possibility of optimistic bias if configuration choices had been tuned on the test set; the text, however, places both choices at dataset construction, and no equation or reported metric reduces to its own inputs. Overall, the derivation chain is self-contained and the circularity score is low (1), reflecting only minor non-load-bearing self-citations.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are all machine learning hyperparameters or probe design choices, not physical constants. The main assumptions are about the fidelity of the turbulence simulation and the relevance of the chosen class labels. The Reynolds number definition is the most ad hoc element.

free parameters (3)

loss weights w1 and w2 = w1=0.7, w2=0.4
Empirically selected from a grid search over weight combinations to maximize median validation accuracy, as shown in Figure 6.
incident OAM topological charge l = l=3
Selected after evaluating l=1 to 5 and observing peak performance at l=3 (Figure 9). This is a post-hoc choice of a probe parameter based on test results.
OAM spectral range = l in [-20, +19] (40 modes)
Adopted after truncation analysis (Figure 10), where reducing the range degraded Model 3 performance, so the full range was chosen to maximize accuracy.

assumptions (5)

domain assumption Kolmogorov turbulence statistics and the Von Karman power spectral density describe atmospheric refractive index fluctuations.
Invoked in Section 2 to generate phase screens; standard but not universally valid for all atmospheric conditions.
domain assumption The split-step Fourier method with 150 phase screens adequately simulates 1500 m propagation of an LG beam.
Section 2 states this approach 'balances numerical efficiency and physical accuracy'; the number of screens and grid resolution are chosen by the authors without convergence analysis.
ad hoc to paper Reynolds number Re=(L0/l0)^(4/3) is a meaningful parameter for atmospheric turbulence classification.
Section 2 defines Re in this way, and the paper uses it as a class label. This relation is not standard in atmospheric optics, where inner scale is typically used directly.
domain assumption The scintillation index model in Eq. (2) applies to LG beams across weak to strong turbulence.
Used to compute Table 1 values; the expression is an analytic model for plane waves and may not accurately describe vortex beams at very high scintillation indices (up to 12.2).
domain assumption Far-field intensity patterns and OAM spectra derived from the same complex field are distinct, complementary modalities.
Stated in Section 1; they are deterministic functions of the same field, so their independence is not established, though they emphasize different physical aspects (amplitude vs phase).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine learning assisted speckle and OAM spectrum analysis for enhanced turbulence characterisation." pith.science (2026). https://pith.science/paper/ITMRIEGK

@misc{pith2026250521878,
  author       = {Pith},
  title        = {Pith review of: Machine learning assisted speckle and OAM spectrum analysis for enhanced turbulence characterisation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITMRIEGK}},
  note         = {Machine review of arXiv:2505.21878}
}

read the original abstract

Atmospheric turbulence degrades the performance of free-space optical (FSO) communication and remote sensing systems by introducing phase and intensity distortions. While a majority of research focuses on mitigating these effects to ensure robust signal transmission, an underexplored alternative is to leverage the transformation of structured light to characterize the turbulent medium itself. Here, we introduce a deep learning framework that fuses post-propagation intensity speckle patterns and orbital angular momentum (OAM) spectral data for atmospheric turbulence parameter inference. Our architecture, based on a modified InceptionNet backbone, is optimized to extract and integrate multi-scale features from these distinct optical modalities. This multimodal approach achieves validation accuracies exceeding 80%, substantially outperforming conventional single-modality baselines. The framework demonstrates high inference accuracy and enhanced training stability across a broad range of simulated turbulent conditions, quantified by varying Fried parameters (r0) and Reynolds numbers (Re). This work presents a scalable and data-efficient method for turbulence characterization, offering a pathway toward robust environmental sensing and the optimization of dynamic FSO systems.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

[1]

Weng and Z

Y. Weng and Z. Pan, Orbital Angular Momentum Based Sensing and Their Applications: A Review, J. Lightwave Technol. PP, 1–10 (2022)

work page 2022
[2]

768-ary Laguerre-Gaussian-mode shift keying free-space optical communication based on convolutional neural networks

Luan H, Lin D, Li K, et al. 768-ary Laguerre-Gaussian-mode shift keying free-space optical communication based on convolutional neural networks. Optics Express, 29, 13, 19807-19818 (2021)

work page 2021
[3]

Bongiovanni, D

D. Bongiovanni, D. Li, M. Goutsoulas et al., Free-space realization of tunable pin-like optical vortex beams, Photon. Res. 9, 1204–1212 (2021)

work page 2021
[4]

Cheng, W

M. Cheng, W. Jiang, L. Guo et al., Metrology with a twist: probing and sensing with vortex light, Light Sci. Appl. 14, 4 (2025)

work page 2025
[5]

Orbital angular momentum-mediated machine learning for high-accuracy mode- feature encoding

Fang X, Hu X, Li B, et al. Orbital angular momentum-mediated machine learning for high-accuracy mode- feature encoding. Light Sci. Appl, 13, 1, 49 (2024)

work page 2024
[6]

Robust occlusion-aware orbital angular momentum feature extraction via all-optical diffractive processing systems

Li K, Jia Y, Gu M, et al. Robust occlusion-aware orbital angular momentum feature extraction via all-optical diffractive processing systems. Optics Express, 33, 11, 23053-23064 (2025)

work page 2025
[7]

Forbes, M

A. Forbes, M. De Oliveira, M. R. Dennis, Structured light, Nat. Photon. 15(4): 253-262 (2021)

work page 2021
[8]

Andrews, L. C. & Beason, M. K. Laser beam propagation in random media: new and advanced topics (2023)

work page 2023

Show all 30 references

[9]

Wang, Advances in communications using optical vortices, Photon

J. Wang, Advances in communications using optical vortices, Photon. Res. 4, B14–B28 (2016)

2016
[10]

A. Klug, C. Peters, and A. Forbes, Robust structured light in atmospheric turbulence, Adv. Photon. 5, 016006 (2023)

2023
[11]

Fu and C

S. Fu and C. Gao, Influences of atmospheric turbulence effects on the orbital angular momentum spectra of vortex beams, Photon. Res. 4, B1–B4 (2016)

2016
[12]

Cheng and M

M. Cheng and M. Lavery, Optical angular momentum interaction with turbulent and scattering media, in Structured Light for Optical Communication (Elsevier), 237–258 (2021)

2021
[13]

Gu and G

Y. Gu and G. Gbur, Measurement of atmospheric turbulence strength by vortex beam, Opt. Commun. 283, 1209–1212 (2010)

2010
[14]

Belmonte, C

A. Belmonte, C. Rosales-Guzmá n, and J. P. Torres, Measurement of flow vorticity with helical beams of light, Optica 2, 1002 (2015)

2015
[15]

Ryabtsev, S

A. Ryabtsev, S. Pouya, A. Safaripour, M. Koochesfahani, and M. Dantus, Fluid flow vorticity measurement using laser beams with orbital angular momentum, Opt. Express 24, 11762–11772 (2016)

2016
[16]

A. H. Dorrah, M. Zamboni-Rached, and M. Mojahedi, Experimental demonstration of tunable refractometer based on orbital angular momentum of longitudinally structured light, Light Sci. Appl. 7, 40 (2018)

2018
[17]

E. F. Strong et al., Angular velocimetry for fluid flows: an optical sensor using structured light and machine learning, Opt. Express 29, 9960–9976 (2021)

2021
[18]

Predicting atmospheric turbulence for secure quantum communications in free space

Jaouni T, Scarfe L, Bouchard F, et al. Predicting atmospheric turbulence for secure quantum communications in free space. Optics Express, 33, 5, 10759-10776 (2025)

2025
[19]

R. D. Sun, L. Guo, M. Cheng et al., Identifying orbital angular momentum modes in turbulence with high accuracy via machine learning, J. Opt. 21, 075703 (2019)

2019
[20]

J. Li, M. Zhang, D. Wang, S. Wu, and Y. Zhan, Joint atmospheric turbulence detection and adaptive demodulation technique using the CNN for the OAM-FSO communication, Opt. Express 26, 10494–10508 (2018)

2018
[21]

W. Yin, B. Zhang, P. Meng et al., A neural network method for inversion of turbulence strength, J. Nonlinear Math. Phys. 31, 22 (2024)

2024
[22]

Yin et al., Experimental study of atmospheric turbulence detection using an orbital angular momentum beam via a convolutional neural network, IEEE Access 7, 184235–184241 (2019)

X. Yin et al., Experimental study of atmospheric turbulence detection using an orbital angular momentum beam via a convolutional neural network, IEEE Access 7, 184235–184241 (2019)

2019
[23]

H. Lv, Y. Guo, Z. X. Yang et al., Identification of diffracted vortex beams at different propagation distances using deep learning, Front. Phys. 10, 843932 (2022)

2022
[24]

P. Meng, J. Zhuang, L. Zhou, W. Yin, and D. Qi, Efficient synchronous retrieval of OAM modes and AT strength using multi-task neural networks, Opt. Express 32, 7816–7831 (2024)

2024
[25]

Zhou et al., Atmospheric turbulence strength distribution along a propagation path probed by longitudinally structured optical beams, Nat

H. Zhou et al., Atmospheric turbulence strength distribution along a propagation path probed by longitudinally structured optical beams, Nat. Commun. 14, 4701 (2023)

2023
[26]

Z. Chen, U. Daly, A. Boldin, and M. P. J. Lavery, Weather sensing with structured light, Commun. Phys. 8, 1 (2025)

2025
[27]

J. Zhao, Z. Zhang, Y. Li, and L. Cen, Multidimensional information assisted deep learning realizing flexible recognition of vortex beam modes, IEEE Photon. J. 13, 1–6 (2021)

2021
[28]

K. Dong, B. Tao, M. Cheng et al., Characterizing propagation and vortex-splitting dynamics of Bessel-Gaussian beams in short-range atmospheric conditions, Opt. Express 33, 2878–2895 (2025)

2025
[29]

Peters, V

C. Peters, V. Cocotos, and A. Forbes, Structured light in atmospheric turbulence—a guide to its digital implementation: tutorial, Adv. Opt. Photon. 17, 113–184 (2025)

2025
[30]

Szegedy, W

C. Szegedy, W. Liu, Y. Jia et al., Going deeper with convolutions, in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. 1–9 (2015)

2015

Pith tools

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

[1] [1]

Weng and Z

Y. Weng and Z. Pan, Orbital Angular Momentum Based Sensing and Their Applications: A Review, J. Lightwave Technol. PP, 1–10 (2022)

work page 2022

[2] [2]

768-ary Laguerre-Gaussian-mode shift keying free-space optical communication based on convolutional neural networks

Luan H, Lin D, Li K, et al. 768-ary Laguerre-Gaussian-mode shift keying free-space optical communication based on convolutional neural networks. Optics Express, 29, 13, 19807-19818 (2021)

work page 2021

[3] [3]

Bongiovanni, D

D. Bongiovanni, D. Li, M. Goutsoulas et al., Free-space realization of tunable pin-like optical vortex beams, Photon. Res. 9, 1204–1212 (2021)

work page 2021

[4] [4]

Cheng, W

M. Cheng, W. Jiang, L. Guo et al., Metrology with a twist: probing and sensing with vortex light, Light Sci. Appl. 14, 4 (2025)

work page 2025

[5] [5]

Orbital angular momentum-mediated machine learning for high-accuracy mode- feature encoding

Fang X, Hu X, Li B, et al. Orbital angular momentum-mediated machine learning for high-accuracy mode- feature encoding. Light Sci. Appl, 13, 1, 49 (2024)

work page 2024

[6] [6]

Robust occlusion-aware orbital angular momentum feature extraction via all-optical diffractive processing systems

Li K, Jia Y, Gu M, et al. Robust occlusion-aware orbital angular momentum feature extraction via all-optical diffractive processing systems. Optics Express, 33, 11, 23053-23064 (2025)

work page 2025

[7] [7]

Forbes, M

A. Forbes, M. De Oliveira, M. R. Dennis, Structured light, Nat. Photon. 15(4): 253-262 (2021)

work page 2021

[8] [8]

Andrews, L. C. & Beason, M. K. Laser beam propagation in random media: new and advanced topics (2023)

work page 2023