Real-time surrogate modeling of nonlinear pulse evolution in multimode fibers

Bahad{\i}r Utku Kesgin; Bora \c{C}arp{\i}nl{\i}o\u{g}lu; U\u{g}ur Te\u{g}in

REVIEW 3 major objections 5 minor 49 references

Real-time surrogate modeling of nonlinear pulse evolution in multimode fibers

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A U-Net surrogate predicts nonlinear multimode fiber propagation at ~88% structural similarity to beam-propagation simulations.

desk verdict Solid short-range forward surrogate with a real temporal-state gap that the paper's generalization claim glosses over. read the letter →

arxiv 2505.12517 v1 pith:65RN6NJ7 submitted 2025-05-18 physics.optics

Bora Çarp{i}nl{i}oğlu , Bahad{i}r Utku Kesgin , Uğur Teğin This is my paper · ORCID

classification physics.optics

keywords multimodefibernonlinearpulsepropagationU-Netsurrogatemodelbeammethodstructuralsimilarityindexspatiotemporalinstabilitymodedecomposition

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 aims to show that a U-Net, a fully convolutional encoder-decoder network, can act as a fast and accurate surrogate for nonlinear pulse propagation in multimode fibers, where direct simulation is expensive. The authors train the network on 4,250 input–output pairs of two-dimensional spatial beam profiles extracted from full spatiotemporal simulations of the generalized nonlinear Schrödinger equation, and report an average structural similarity of about 0.88 with beam-propagation results on held-out test fields. They further show that the same network can be cascaded to predict propagation beyond the trained distance, and that reversing the input and output roles learns an approximate inverse model that estimates input fields from desired outputs. If these results hold, the approach offers near-real-time surrogate simulation and a practical route toward tasks such as mode decomposition and targeted excitation of nonlinear effects.

What carries the argument

The central mechanism is the U-Net, an encoder–decoder convolutional network whose skip connections preserve fine spatial detail while the bottleneck compresses global context; for this task it maps one two-dimensional complex optical field to another by encoding amplitude and phase (or real and imaginary parts) as two input and two output channels. The ground truth is generated by numerically integrating the generalized nonlinear Schrödinger equation with the symmetrized split-step Fourier method, which supplies the training pairs. Generalization to untrained distances is achieved by cascading the trained network, using each predicted field as the next input, which works because the network preserves and passes through both complex-field channels.

What would settle it

Propagate several input fields that have identical spatial beam profiles but different temporal envelopes (for example, different pulse durations or chirps) to 2.5 cm with the beam propagation method, and compare the resulting output spatial profiles with the U-Net's single prediction; if the outputs differ substantially and the network cannot match them at SSIM near 0.88, the learned map is not a valid surrogate outside its training distribution.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a U-Net trained on simulated spatial field pairs learns the input-to-output spatial map of nonlinear propagation through a graded-index multimode fiber at a fixed distance and power. The training data come from split-step Fourier beam-propagation simulations of spatiotemporal instability, with each input built from five randomly chosen modes out of 120 propagated over 2.5 cm at 5 MW peak power. The network treats the complex field as two channels and reaches average SSIM of 0.8823 for amplitude and 0.8909 for phase on the test set, with mean squared error around 0.1%. Cascading the same network produces predictions at 2.5 cm intervals that grow less accurate with distance but remain far faster than direct simulation. Reversing the dataset yields an inverse model with average SSIM of 0.6998 (amplitude) and 0.7828 (phase), which the authors present as promising for mode decomposition and inverse design.

Load-bearing premise

The two-dimensional spatial profile of the input beam fully determines the two-dimensional spatial profile of the output beam, so the temporal shape of the pulse can be ignored even though the governing equation couples spatial and temporal dynamics through dispersion and nonlinearity.

Editorial extensions

If this is right

A fast surrogate for short-distance multimode fiber propagation becomes available, enabling real-time iterative optimization loops that direct beam-propagation solvers cannot support.
Cascading the same network gives a tunable speed–accuracy trade-off for extended propagation distances, useful when approximate predictions at 5–15 cm are sufficient.
The inverse-trained U-Net can estimate input spatial fields for a desired output, which can support mode decomposition and the targeted excitation of specific nonlinear phenomena.
The same training recipe—simulated field pairs feeding a U-Net with complex fields as two channels—can be applied to other nonlinear photonic components where simulation is the bottleneck.

Reading between the lines

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

Because the dataset discards the temporal envelope, the surrogate is only validated for pulses with the same temporal structure as the training set; a test with different pulse durations or chirps would reveal whether the spatial mapping is well-posed beyond that distribution.
The cascade's gradual SSIM drop suggests the learned map approximates the propagator only on the training manifold; training on pairs sampled at multiple distances could reduce error accumulation and is a natural extension the authors do not test.
The same architecture could be extended to full spatiotemporal prediction by treating time as an additional channel or using spatiotemporal convolutions, an extension the authors mention but do not implement.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things to know. The forward U-Net at the trained distance is credible: ~0.88 SSIM on a held-out test set, clean dataset generation, standard train/test split. But the headline claim that it generalizes to untrained propagation distances is not supported as written.

The load-bearing problem is that the model discards the temporal dimension. The GNLSE in Eq. (1) has dispersion and nonlinearity acting on the full spatiotemporal field, yet the dataset is built from pairs of 2D spatial profiles. Unless the temporal envelope was fixed for all 4250 examples — the paper never says — the mapping is under-determined even at 2.5 cm. When the trained U-Net is cascaded, each step feeds it a spatial profile that was produced by a temporal envelope different from the training one. The model has no way to know that, so the error compounds. Figure 3 shows intensity SSIM falling to 0.38 at 5 cm and 0.17 at 15 cm. That is not 'generalization'; it is a short-range interpolant with predictable drift.

Credit where due: this is a new application of a known architecture, done carefully for the narrow regime. The 2.5 cm forward result is a legitimate proof of concept, with amplitude and phase SSIM both near 0.89. The inverse model is honestly presented as a separate network with lower performance. The authors also state upfront that there is no experimental validation.

Soft spots beyond the temporal issue: no error bars on the SSIM figures, no quantitative comparison against the RNN baselines they claim to be slower (the supplement might have it, but the main text does not), and no public code or data. That makes reproducibility hard to assess.

Who this is for: anyone working on fast design loops or control in multimode fibers at fixed input power and fixed fiber length, where a 2.5 cm step surrogate could be useful.

Recommendation: I would send this to peer review — the protocol is sound enough that referees can check it — but I would expect a major revision. The authors need to either restrict the claim to a fixed temporal envelope or include temporal information in the model, and the abstract should not say 'generalize to untrained propagation distances' without a qualification. If code and data are released, I might cite the short-range result. As it stands, I wouldn't rely on the cascade.

Referee Report

3 major / 5 minor

Summary. The paper proposes a U-Net surrogate for nonlinear pulse propagation in multimode fibers. The authors generate 4,250 input-output pairs by beam propagation method (BPM) simulations of the generalized nonlinear Schrödinger equation, extract two-dimensional spatial beam profiles from the full spatiotemporal fields, and train a U-Net to map input spatial profiles to output spatial profiles after 2.5 cm propagation. They report an average SSIM of about 0.88 on a held-out test set. They then apply the trained network iteratively to simulate propagation to longer distances, report a separate U-Net for the inverse problem of predicting input fields from desired outputs, and claim that the model generalizes to untrained propagation distances.

Significance. If the short-distance forward result is taken at face value, the paper demonstrates a fast spatial-profile surrogate for a specific BPM configuration, with a clear training and evaluation protocol and a held-out test set. This could be useful for real-time spatial-profile prediction at a fixed propagation distance and fixed pulse parameters. However, the broader significance is limited by two issues: the learned map discards the temporal degree of freedom that the governing equation depends on, and the reported cascade results show rapid degradation, so the generalization claim in the abstract is not supported by the evidence. The inverse-problem demonstration is suggestive but lacks baselines and error analysis.

major comments (3)

[Dataset generation, Eq. (1)] The mapping learned by the U-Net is from two-dimensional spatial profiles to spatial profiles, but the governing equation (1) contains temporal derivatives (beta2 and beta3 terms) and the nonlinear term gamma|A|^2A, so the output spatial profile depends on the full spatiotemporal input A(x,y,0,t), not only on the spatial profile A(x,y,0). The paper does not state whether the temporal envelope is identical for all 4,250 training samples or how it is varied. If it is fixed, the surrogate is valid only for that one pulse shape; if it varies, many distinct spatiotemporal inputs collapse to the same spatial input, making the learned function ill-posed. This restriction must be stated explicitly, and all claims about generalization must be scoped accordingly.
[Results, Fig. 3] The support for the abstract claim that "the model can generalize to untrained propagation distances" is a single cascade example whose SSIM falls from 0.907 at 2.5 cm to 0.377 at 5 cm, 0.283 at 7.5 cm, and 0.171 at 15 cm. These values indicate that the cascade essentially fails beyond one or two steps. No aggregate statistics, error bars, or comparisons with alternative fast approximations are provided for the cascaded predictions, so the conclusion that the model "generalizes to longer propagation distances" overstates the evidence. The authors should either report cascade performance over the full test set with error bars or revise the claim to say that the model can be iterated with rapidly degrading accuracy.
[Inverse problem, Fig. 4] The inverse U-Net's average SSIM values (0.70 for amplitude, 0.78 for phase) are reported without any comparison to existing mode decomposition or wavefront-shaping baselines, and the predicted fields are not checked for physical consistency such as power conservation or modal content. Since the forward model itself depends on the temporal envelope, the inverse map inherits the same ambiguity. The statement that this result "pave[s] the way for future studies on specific spatial modes contributing to nonlinear phenomena" is therefore not yet supported; at minimum, an aggregate error analysis and a comparison to a simple linear reconstruction baseline are needed.

minor comments (5)

[Figures 2 and 4] The scale bars labeled "5 m" in the figure panels are almost certainly micrometers; please verify the units in the figures and captions.
[Supplement 1 references] Fiber parameters, architecture details, hyperparameter search results, and the claimed inference-time advantage over RNNs are deferred to Supplement 1, but the supplement is not included with the manuscript. Please include the supplement or summarize the key numbers, especially the speed comparison, in the main text.
[Results section] Reported SSIM values are averages, but no standard deviations, test-set sizes, or per-sample distributions are given for the forward or inverse tasks; adding these would help the reader judge the spread across the 20% test split.
[Dataset generation] The phrase "equal linear combination of five spatial modes randomly selected from a total of 120 modes" should specify whether the coefficients are complex with random phases and whether the temporal envelope is the same for every sample.
[Data availability] The data availability statement says the data are not publicly available, and no code is provided; for a machine-learning study, releasing the trained model and dataset-generation code would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the U-Net surrogate is trained and evaluated against held-out BPM simulations, so the reported SSIM and cascade results are empirical, not tautological.

full rationale

The paper claims that a U-Net trained on beam-propagation-method data accurately maps input spatial beam profiles to output spatial beam profiles after 2.5 cm of nonlinear multimode-fiber propagation, and that cascading this network can be applied to longer distances. I find no circular step in this chain. The physical GNLSE in Eq. (1) is used only to generate the training dataset via SSFM/BPM; the U-Net itself is a learned image-to-image regression and does not encode or reuse the BPM solver at inference. The central quantitative claims—0.8823 average SSIM for amplitude, 0.8909 for phase on the forward task, and the degrading cascade SSIM values in Fig. 3—are measured on held-out test samples against independent BPM ground-truth simulations, so they are empirical benchmark results rather than quantities recovered by construction. The self-citations (refs. 5, 6, 11, 14, 17, 20) appear only in contextual passages about prior neural-network approaches and spatiotemporal instability; none is used to justify a load-bearing assumption or to forbid alternative architectures. The temporal-envelope issue raised by the reader is a legitimate concern about physical validity and generalization outside the training distribution, but it is not circularity: the learned mapping is well-defined for the fixed pulse parameters used in dataset generation, and the cascade failure is an empirical limitation explicitly documented by the decreasing SSIM values. No equation-level definitional reduction, fitted parameter renamed as prediction, or self-citation chain is present, so the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The surrogate itself is a fitted model; the only fitted quantities are the network weights and hyperparameters. The governing equation and solver are taken from prior literature. No new physical entities are introduced.

free parameters (1)

U-Net hyperparameters = Not disclosed (optimized via Ray Tune)
Accuracy of the surrogate depends on architecture depth, channel counts, learning rate, and other training choices. The main text reports no values, so replication requires guessing or contacting the authors.

assumptions (4)

domain assumption Generalized nonlinear Schrödinger equation (Eq. 1) governs multimode pulse propagation
The BPM data used as ground truth are generated by solving this equation; if this model is wrong, the surrogate learns the wrong physics.
domain assumption Split-step Fourier method / BPM provides accurate numerical ground truth
Dataset generation uses SSFM; no convergence study is reported to show the numerical error is negligible relative to the reported SSIM values.
domain assumption A U-Net with complex fields encoded as two real channels can approximate the spatial-profile mapping
The universal function approximation of CNNs is assumed; the paper does not analyze capacity or expressiveness for this specific mapping.
ad hoc to paper The 4,250 training pairs are representative of the target regime
Inputs are equal combinations of five randomly chosen modes out of 120, all at 5 MW peak power. The surrogate is only expected to work within this narrow distribution, and no out-of-distribution test beyond distance is shown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Real-time surrogate modeling of nonlinear pulse evolution in multimode fibers." pith.science (2026). https://pith.science/paper/65RN6NJ7

@misc{pith2026250512517,
  author       = {Pith},
  title        = {Pith review of: Real-time surrogate modeling of nonlinear pulse evolution in multimode fibers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65RN6NJ7}},
  note         = {Machine review of arXiv:2505.12517}
}

abstract

Modeling nonlinear pulse propagation in multimode fibers is challenging due to the large number of interacting modes and the resulting spatiotemporal complexity. Traditional optimization methods often become intractable, while learning-based approaches, such as recurrent neural networks, suffer from high computational cost and long inference times. We present a U-Net architecture as a fast, accurate surrogate for modeling nonlinear pulse propagation in multimode fibers. This approach overcomes the intractability of traditional methods while offering low computational cost. Trained on data generated by beam propagation method, our approach achieves an $\sim$88\% average structural similarity index with simulations. The model can generalize to untrained propagation distances, demonstrating convolutional architectures as efficient tools for simulating complex spatiotemporal dynamics in multimode fibers and offering potential for applications like mode decomposition.

Figures

Figures reproduced from arXiv: 2505.12517 by the authors.

**Figure 1.** The overall scheme of dataset generation and fitting a U-Net model to the generated dataset. A multimode fiber (MMF) is used to map input spatial patterns consisting of mode combinations to output spatial patterns. promising avenue for applications such as mode decompositionbased analyses and tailored pulse shaping. Dataset generation To generate a dataset exhibiting the rich spatiotemporal dynamics essential for t… view at source ↗

**Figure 2.** U-Net training results for 2.5 cm propagation in a multimode fiber. The amplitude (top) and phase (bottom) parts of predicted and ground truth complex fields are compared together with corresponding structural similarity index values [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Normalized intensity profiles obtained from BPM simulations (left) and by cascading the trained U-Nets (right) six times illustrated at integer multiples of the initial propagation distance, 2.5 cm. SSIM values are given to the left of U-Net predictions. our specific task, we leveraged the automatic hyperparameter search capabilities of the Ray Tune library [22] for each training scenario explored in this study. Re… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: U-Net training results for the inverse problem where the roles of the input and output are reversed. The amplitude (top) and phase (bottom) parts of predicted and ground truth complex fields are compared together with corresponding structural similarity index values. C…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 46 canonical work pages

[1]

Skarsoulis, K

K. Skarsoulis, K. Makris, C. Moser, and D. Psaltis, Optica11, 782 (2024)

work page 2024
[2]

Oguz, J.-L

I. Oguz, J.-L. Hsieh, N. U. Dinc,et al., Adv. Photonics6, 016002 (2024)

work page 2024
[4]

L. G. Wright, Z. M. Ziegler, P . M. Lushnikov,et al., IEEE J. Sel. Top. Quantum Electron.24, 1 (2017)

work page 2017
[5]

B. U. Kesgin and U. Te ˘gin, Nanophotonics (2025)

work page 2025
[6]

A. M. Iskandar Muda and U. Te ˘gin, Opt. Express33, 7852 (2025)

work page 2025
[8]

H. Cao, T. ˇCižmár, S. Turtaev,et al., Adv. Opt. Photonics15, 524 (2023). Publisher: Optica Publishing Group

work page 2023
[9]

Wisal, C.-W

K. Wisal, C.-W. Chen, Z. Kuang,et al., Optica11, 1663 (2024). Pub- lisher: Optica Publishing Group

work page 2024
[10]

Tzang, A

O. Tzang, A. M. Caravaca-Aguirre, K. Wagner, and R. Piestun, Nat. Photonics12, 368 (2018)

work page 2018

Show all 49 references

[11]

Te˘gin, B

U. Te˘gin, B. Rahmani, E. Kakkava,et al., Adv. Photonics2, 056005 (2020)

2020
[12]

Freire, E

P . Freire, E. Manuylovich, J. E. Prilepsky, and S. K. Turitsyn, Adv. Opt. Photonics15, 739 (2023). Publisher: Optica Publishing Group

2023
[13]

Genty, L

G. Genty, L. Salmela, J. M. Dudley,et al., Nat. Photonics15, 91 (2021). Publisher: Nature Publishing Group UK London

2021
[14]

Te˘gin, B

U. Te˘gin, B. Rahmani, E. Kakkava,et al., Apl Photonics5(2020). Publisher: AIP Publishing

2020
[15]

Xiong, B

W. Xiong, B. Redding, S. Gertler,et al., APL Photonics5(2020). Publisher: AIP Publishing

2020
[16]

Salmela, N

L. Salmela, N. Tsipinakis, A. Foi,et al., Nat. machine intelligence3, 344 (2021). Publisher: Nature Publishing Group UK London

2021
[17]

Te˘gin, N

U. Te˘gin, N. U. Dinç, C. Moser, and D. Psaltis, Nat. Mach. Intell.3, 387 (2021). Publisher: Nature Publishing Group UK London

2021
[18]

Jiang and Y .-C

J. Jiang and Y .-C. Lai, Phys. review research1, 033056 (2019). Pub- lisher: APS

2019
[19]

Krupa, A

K. Krupa, A. Tonello, A. Barthélémy,et al., Phys. review letters116, 183901 (2016). Publisher: APS

2016
[20]

Te˘gin and B

U. Te˘gin and B. Ortaç, IEEE Photonics Technol. Lett.29, 2195 (2017)

2017
[23]

P . S. Anisimov, E. D. Tsyplakov, V. V. Zemlyakov, and J. Gao, Opt. Lett. 48, 2022 (2023)

2023
[24]

Rothe, Q

S. Rothe, Q. Zhang, N. Koukourakis, and J. W. Czarske, Appl. Sci.10, 1367 (2020)

2020
[25]

AI for nonlinear control in fiber-optics,

G. Genty, “AI for nonlinear control in fiber-optics,” inAI and Optical Data Sciences VI,, vol. PC13375 M. Notomi and T. Zhou, eds., International Society for Optics and Photonics (SPIE, 2025), p. PC1337519

2025
[27]

Molesky, Z

S. Molesky, Z. Lin, A. Y . Piggott,et al., Nat. Photonics12, 659 (2018). Letter 5 FULL REFERENCES

2018
[28]

Ptychographic imaging with a fiber endoscope via wavelength scanning,

K. Skarsoulis, K. Makris, C. Moser, and D. Psaltis, “Ptychographic imaging with a fiber endoscope via wavelength scanning,” Optica11, 782–790 (2024)

2024
[29]

Programming nonlinear propaga- tion for efficient optical learning machines,

I. Oguz, J.-L. Hsieh, N. U. Dinc,et al., “Programming nonlinear propaga- tion for efficient optical learning machines,” Adv. Photonics6, 016002– 016002 (2024)

2024
[30]

Miller,Optical fiber telecommunications(Elsevier, 2012)

S. Miller,Optical fiber telecommunications(Elsevier, 2012)

2012
[31]

Multimode nonlinear fiber optics: massively parallel numerical solver, tutorial, and outlook,

L. G. Wright, Z. M. Ziegler, P . M. Lushnikov,et al., “Multimode nonlinear fiber optics: massively parallel numerical solver, tutorial, and outlook,” IEEE J. Sel. Top. Quantum Electron.24, 1–16 (2017)

2017
[32]

Photonic neural networks at the edge of spatiotemporal chaos in multimode fibers,

B. U. Kesgin and U. Te˘gin, “Photonic neural networks at the edge of spatiotemporal chaos in multimode fibers,” Nanophotonics (2025)

2025
[33]

Optical computing with supercontin- uum generation in photonic crystal fibers,

A. M. Iskandar Muda and U. Te˘gin, “Optical computing with supercontin- uum generation in photonic crystal fibers,” Opt. Express33, 7852–7861 (2025)

2025
[34]

Optical computing using telecom device platform for nonlinear mapping to high- dimensional feature space,

E. Manuylovich, D. Stoliarov, D. Saad, and S. K. Turitsyn, “Optical computing using telecom device platform for nonlinear mapping to high- dimensional feature space,” arXiv preprint arXiv:2501.03429 (2025)

2025 arXiv
[35]

Controlling light propagation in multimode fibers for imaging, spectroscopy, and beyond,

H. Cao, T. ˇCižmár, S. Turtaev,et al., “Controlling light propagation in multimode fibers for imaging, spectroscopy, and beyond,” Adv. Opt. Photonics15, 524–612 (2023). Publisher: Optica Publishing Group

2023
[36]

Optimal input excitations for suppressing nonlinear instabilities in multimode fibers,

K. Wisal, C.-W. Chen, Z. Kuang,et al., “Optimal input excitations for suppressing nonlinear instabilities in multimode fibers,” Optica11, 1663–1672 (2024). Publisher: Optica Publishing Group

2024
[37]

Adap- tive wavefront shaping for controlling nonlinear multimode interactions in optical fibres,

O. Tzang, A. M. Caravaca-Aguirre, K. Wagner, and R. Piestun, “Adap- tive wavefront shaping for controlling nonlinear multimode interactions in optical fibres,” Nat. Photonics12, 368–374 (2018)

2018
[38]

Single-mode output by con- trolling the spatiotemporal nonlinearities in mode-locked femtosecond multimode fiber lasers,

U. Te˘gin, B. Rahmani, E. Kakkava,et al., “Single-mode output by con- trolling the spatiotemporal nonlinearities in mode-locked femtosecond multimode fiber lasers,” Adv. Photonics2, 056005–056005 (2020)

2020
[39]

Arti- ficial neural networks for photonic applications—from algorithms to implementation: tutorial,

P . Freire, E. Manuylovich, J. E. Prilepsky, and S. K. Turitsyn, “Arti- ficial neural networks for photonic applications—from algorithms to implementation: tutorial,” Adv. Opt. Photonics15, 739–834 (2023). Publisher: Optica Publishing Group

2023
[40]

Machine learning and applications in ultrafast photonics,

G. Genty, L. Salmela, J. M. Dudley,et al., “Machine learning and applications in ultrafast photonics,” Nat. Photonics15, 91–101 (2021). Publisher: Nature Publishing Group UK London

2021
[41]

Controlling spatiotemporal nonlinearities in multimode fibers with deep neural networks,

U. Te˘gin, B. Rahmani, E. Kakkava,et al., “Controlling spatiotemporal nonlinearities in multimode fibers with deep neural networks,” Apl Photonics5(2020). Publisher: AIP Publishing

2020
[42]

Deep learning of ultrafast pulses with a multimode fiber,

W. Xiong, B. Redding, S. Gertler,et al., “Deep learning of ultrafast pulses with a multimode fiber,” APL Photonics5(2020). Publisher: AIP Publishing

2020
[43]

Predicting ultrafast nonlinear dynamics in fibre optics with a recurrent neural network,

L. Salmela, N. Tsipinakis, A. Foi,et al., “Predicting ultrafast nonlinear dynamics in fibre optics with a recurrent neural network,” Nat. machine intelligence3, 344–354 (2021). Publisher: Nature Publishing Group UK London

2021
[44]

Reusability report: Predicting spatiotemporal nonlinear dynamics in multimode fibre optics with a recurrent neural network,

U. Te˘gin, N. U. Dinç, C. Moser, and D. Psaltis, “Reusability report: Predicting spatiotemporal nonlinear dynamics in multimode fibre optics with a recurrent neural network,” Nat. Mach. Intell.3, 387–391 (2021). Publisher: Nature Publishing Group UK London

2021
[45]

Model-free prediction of spatiotemporal dynami- cal systems with recurrent neural networks: Role of network spectral radius,

J. Jiang and Y .-C. Lai, “Model-free prediction of spatiotemporal dynami- cal systems with recurrent neural networks: Role of network spectral radius,” Phys. review research1, 033056 (2019). Publisher: APS

2019
[46]

Observation of geometric parametric instability induced by the periodic spatial self-imaging of multimode waves,

K. Krupa, A. Tonello, A. Barthélémy,et al., “Observation of geometric parametric instability induced by the periodic spatial self-imaging of multimode waves,” Phys. review letters116, 183901 (2016). Publisher: APS

2016
[47]

Spatiotemporal instability of femtosecond pulses in graded-index multimode fibers,

U. Te˘gin and B. Ortaç, “Spatiotemporal instability of femtosecond pulses in graded-index multimode fibers,” IEEE Photonics Technol. Lett.29, 2195–2198 (2017)

2017
[48]

U-net: Convolutional net- works for biomedical image segmentation,

O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolutional net- works for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18,...

2015
[49]

Tune: A research platform for dis- tributed model selection and training,

R. Liaw, E. Liang, R. Nishihara,et al., “Tune: A research platform for dis- tributed model selection and training,” arXiv preprint arXiv:1807.05118 (2018)

2018 arXiv
[50]

On similarity metrics evaluating the performance of mode decomposition in few-mode optical fibers,

P . S. Anisimov, E. D. Tsyplakov, V. V. Zemlyakov, and J. Gao, “On similarity metrics evaluating the performance of mode decomposition in few-mode optical fibers,” Opt. Lett.48, 2022–2025 (2023)

2023
[51]

Deep learning for computational mode decomposition in optical fibers,

S. Rothe, Q. Zhang, N. Koukourakis, and J. W. Czarske, “Deep learning for computational mode decomposition in optical fibers,” Appl. Sci.10, 1367 (2020)

2020
[52]

AI for nonlinear control in fiber-optics,

G. Genty, “AI for nonlinear control in fiber-optics,” inAI and Optical Data Sciences VI,vol. PC13375 M. Notomi and T. Zhou, eds., International Society for Optics and Photonics (SPIE, 2025), p. PC1337519

2025
[53]

A Closer Look at Spatiotemporal Convolutions for Action Recognition,

D. Tran, H. Wang, L. Torresani,et al., “A Closer Look at Spatiotemporal Convolutions for Action Recognition,” (2018). _eprint: 1711.11248

2018 arXiv
[54]

Inverse design in nanophoton- ics,

S. Molesky, Z. Lin, A. Y . Piggott,et al., “Inverse design in nanophoton- ics,” Nat. Photonics12, 659–670 (2018)

2018

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Skarsoulis, K

K. Skarsoulis, K. Makris, C. Moser, and D. Psaltis, Optica11, 782 (2024)

work page 2024

[2] [2]

Oguz, J.-L

I. Oguz, J.-L. Hsieh, N. U. Dinc,et al., Adv. Photonics6, 016002 (2024)

work page 2024

[3] [4]

L. G. Wright, Z. M. Ziegler, P . M. Lushnikov,et al., IEEE J. Sel. Top. Quantum Electron.24, 1 (2017)

work page 2017

[4] [5]

B. U. Kesgin and U. Te ˘gin, Nanophotonics (2025)

work page 2025

[5] [6]

A. M. Iskandar Muda and U. Te ˘gin, Opt. Express33, 7852 (2025)

work page 2025

[6] [8]

H. Cao, T. ˇCižmár, S. Turtaev,et al., Adv. Opt. Photonics15, 524 (2023). Publisher: Optica Publishing Group

work page 2023

[7] [9]

Wisal, C.-W

K. Wisal, C.-W. Chen, Z. Kuang,et al., Optica11, 1663 (2024). Pub- lisher: Optica Publishing Group

work page 2024

[8] [10]

Tzang, A

O. Tzang, A. M. Caravaca-Aguirre, K. Wagner, and R. Piestun, Nat. Photonics12, 368 (2018)

work page 2018