Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

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

Matters Arising to this article was published on 13 May 2021

A preprint version of the article is available at arXiv.

Abstract

The propagation of ultrashort pulses in optical fibre plays a central role in the development of light sources and photonic technologies, with applications from fundamental studies of light–matter interactions to high-resolution imaging and remote sensing. However, short pulse dynamics are highly nonlinear, and optimizing pulse propagation for application purposes requires extensive and computationally demanding numerical simulations. This creates a severe bottleneck in designing and optimizing experiments in real time. Here, we present a solution to this problem using a recurrent neural network to model and predict complex nonlinear propagation in optical fibre, solely from the input pulse intensity profile. We highlight particular examples in pulse compression and ultra-broadband supercontinuum generation, and compare neural network predictions with experimental data. We also show how the approach can be generalized to model other propagation scenarios for a wider range of input conditions and fibre systems, including multimode propagation. These results open up novel perspectives in the modelling of nonlinear systems, for the development of future photonic technologies and more generally in physics for studies in Bose–Einstein condensates, plasma physics and hydrodynamics.

Access options

Rent or Buy article

Get time limited or full article access on ReadCube.

from$8.99

All prices are NET prices.

Fig. 1: Recurrent neural networks.
Fig. 2: Temporal-evolution RNN modelling of higher-order soliton propagation dynamics.
Fig. 3: Experimental validation of higher-order soliton temporal dynamics RNN modelling.
Fig. 4: Spectral evolution RNN modelling of higher-order soliton propagation dynamics.
Fig. 5: Experimental validation of higher-order soliton spectral dynamics RNN modelling.
Fig. 6: Temporal and spectral evolution RNN modelling of a supercontinuum.

Data availability

The numerical data used in this work and a public version of the codes are available at https://gitlab.com/salmelal/rnnnonlinear and with full simulation datasets at https://doi.org/10.5281/zenodo.4304771 under an MIT licence.

References

  1. 1.

    Jordan, M. I. & Mitchell, T. M. Machine learning: trends, perspectives and prospects. Science 349, 255–260 (2015).

    MathSciNet  MATH  Article  Google Scholar 

  2. 2.

    Wetzel, B. et al. Customizing supercontinuum generation via on-chip adaptive temporal pulse-splitting. Nat. Commun. 9, 4884 (2018).

    Article  Google Scholar 

  3. 3.

    Michaeli, L. & Bahabad, A. Genetic algorithm driven spectral shaping of supercontinuum radiation in a photonic crystal fiber. J. Opt. 20, 055501 (2018).

    Article  Google Scholar 

  4. 4.

    Tzang, O., Caravaca-Aguirre, A. M., Wagner, K. & Piestun, R. Adaptive wavefront shaping for controlling nonlinear multimode interactions in optical fibres. Nat. Photon. 12, 368–374 (2018).

    Article  Google Scholar 

  5. 5.

    Teğin, U. et al. Controlling spatiotemporal nonlinearities in multimode fibers with deep neural networks. APL Photonics 5, 030804 (2020).

    Article  Google Scholar 

  6. 6.

    Andral, U. et al. Fiber laser mode locked through an evolutionary algorithm. Optica 2, 275–278 (2015).

    Article  Google Scholar 

  7. 7.

    Pu, G., Yi, L., Zhang, L. & Hu, W. Intelligent programmable mode-locked fiber laser with a human-like algorithm. Optica 6, 362–369 (2019).

    Article  Google Scholar 

  8. 8.

    Dudley, J. M. & Meng, F. Toward a self-driving ultrafast fiber laser. Light Sci. Appl. 9, 26 (2020).

    Article  Google Scholar 

  9. 9.

    Kokhanovskiy, A., Ivanenko, A., Kobtsev, S., Smirnov, S. & Turitsyn, S. Machine learning methods for control of fibre lasers with double gain nonlinear loop mirror. Sci. Rep. 9, 2916 (2019).

    Article  Google Scholar 

  10. 10.

    Kokhanovskiy, A. et al. Machine learning-based pulse characterization in figure-eight mode-locked lasers. Opt. Lett. 44, 3410–3413 (2019).

    Article  Google Scholar 

  11. 11.

    Zahavy, T. et al. Deep learning reconstruction of ultrashort pulses. Optica 5, 666–673 (2018).

    Article  Google Scholar 

  12. 12.

    Närhi, M. et al. Machine learning analysis of extreme events in optical fibre modulation instability. Nat. Commun. 9, 4923 (2018).

    Article  Google Scholar 

  13. 13.

    Baumeister, T., Brunton, S. L. & Kutz, J. N. Deep learning and model predictive control for self-tuning mode-locked lasers. J. Opt. Soc. Am. B 35, 617–626 (2018).

    Article  Google Scholar 

  14. 14.

    Finot, C., Gukov, I., Hammani, K. & Boscolo, S. Nonlinear sculpturing of optical pulses with normally dispersive fiber-based devices. Opt. Fiber Technol. 45, 306–312 (2018).

    Article  Google Scholar 

  15. 15.

    Brunton, S. L., Proctor, J. L. & Kutz, J. N. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl Acad. Sci. USA 113, 3932–3937 (2016).

    MathSciNet  MATH  Article  Google Scholar 

  16. 16.

    Raissi, M. Deep hidden physics models: deep learning of nonlinear partial differential equations. J. Mach. Learn. Res. 19, 932–955 (2018).

    MathSciNet  MATH  Google Scholar 

  17. 17.

    Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707 (2019).

    MathSciNet  MATH  Article  Google Scholar 

  18. 18.

    Vlachas, P. R., Byeon, W., Wan, Z. Y., Sapsis, T. P. & Koumoutsakos, P. Data-driven forecasting of high-dimensional chaotic systems with long short-term memory networks. Proc. R. Soc. A Math. Phys. Eng. Sci. 474, 20170844 (2018).

    MathSciNet  MATH  Google Scholar 

  19. 19.

    Vlachas, P. R. et al. Backpropagation algorithms and reservoir computing in recurrent neural networks for the forecasting of complex spatiotemporal dynamics. Neural Netw. 126, 191–217 (2020).

    Article  Google Scholar 

  20. 20.

    Pandey, S. & Schumacher, J. Reservoir computing model of two-dimensional turbulent convection. Phys. Rev. Fluids 5, 113506 (2020).

    Article  Google Scholar 

  21. 21.

    Jiang, J. & Lai, Y.-C. Model-free prediction of spatiotemporal dynamical systems with recurrent neural networks: role of network spectral radius. Phys. Rev. Res. 1, 033056 (2019).

    Article  Google Scholar 

  22. 22.

    Tikan, A. et al. Universality of the Peregrine soliton in the focusing dynamics of the cubic nonlinear Schrödinger equation. Phys. Rev. Lett. 119, 033901 (2017).

    Article  Google Scholar 

  23. 23.

    Agrawal, G. P. Nonlinear Fiber Optics 5th edn (Academic Press, 2013).

  24. 24.

    Lipton, Z. C. A critical review of recurrent neural networks for sequence learning. Preprint at https://arxiv.org/pdf/1506.00019.pdf (2015).

  25. 25.

    Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput. 9, 1735–1780 (1997).

    Article  Google Scholar 

  26. 26.

    Carleo, G. et al. Machine learning and the physical sciences. Rev. Mod. Phys. 91, 045002 (2019).

    Article  Google Scholar 

  27. 27.

    Yu, Y., Si, X., Hu, C. & Zhang, J. A review of recurrent neural networks: LSTM cells and network architectures. Neural Comput. 31, 1–36 (2019).

    MathSciNet  MATH  Article  Google Scholar 

  28. 28.

    Dudley, J. M., Genty, G. & Coen, S. Supercontinuum generation in photonic crystal fiber. Rev. Mod. Phys. 78, 1135 (2006).

    Article  Google Scholar 

  29. 29.

    Farfan, C. A., Epstein, J. & Turner, D. B. Femtosecond pulse compression using a neural-network algorithm. Opt. Lett. 43, 5166–5169 (2018).

    Article  Google Scholar 

  30. 30.

    Wright, L. G. et al. Multimode nonlinear fiber optics: massively parallel numerical solver, tutorial and outlook. IEEE J. Select. Top. Quantum Electron. 24, 1–16 (2017).

    Article  Google Scholar 

  31. 31.

    Werbos, P. J. Backpropagation through time: what it does and how to do it. Proc. IEEE 78, 1550–1560 (1990).

    Article  Google Scholar 

  32. 32.

    Chollet, F. et al. Keras https://keras.io (2015).

  33. 33.

    Abadi, M. et al. Tensorflow: a system for large-scale machine learning. In Proc. 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16) 265–283 (UESNIX Association, 2016).

  34. 34.

    Goodfellow, I et al. Deep Learning (MIT Press, 2016).

Download references

Acknowledgements

L.S. acknowledges support from the Faculty of Engineering and Natural Sciences Graduate School of Tampere University. J.M.D. acknowledges support from the French Agence Nationale de la Recherche (ANR-20-CE30-0004, ANR-15-IDEX-0003 and ANR-17-EURE-0002). G.G. acknowledges support from the Academy of Finland (298463, 318082 and Flagship PREIN 320165). We also thank D. Brunner for useful discussions.

Author information

Affiliations

Authors

Contributions

L.S. performed all the numerical simulations. All authors participated in the data analysis reported and in the writing of the manuscript. L.S., J.M.D. and G.G. planned the research project and G.G. provided overall supervision.

Corresponding author

Correspondence to Goëry Genty.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Peer review information Nature Machine Intelligence thanks the anonymous reviewers for their contribution to the peer review of this work.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended Data Fig. 1 Comparison of spectral intensity evolution modelling by the RNN when trained using data input in linear and logarithmic (dB) representation.

The dashed red line shows the simulated spectra of a higher-order soliton (N = 6) at selected distances along the fibre similar to those plotted in Fig. 5. The solid blue lines shows the RNN prediction when trained directly with spectral data input in dB. The solid green line shows the RNN prediction (plotted in dB) when trained with spectral data input in linear scale.

Extended Data Fig. 2 Generalization of RNN applicability to modelling NLSE dynamics using normalized training simulations.

Training data were generated using the normalized form of the NLSE (see Methods). The results are plotted in dimensional units. (a) and (b) shows the temporal and spectral evolution corresponding to the initial stage of higher-order soliton compression and breakup with soliton number N = 4, pulse duration (FWHM) of 300 fs and peak power of 157 W. (c) and (d) shows the temporal and spectral evolution corresponding to the initial stage of higher-order soliton compression and breakup with soliton number N = 7, pulse duration (FWHM) of 3 fs and peak power of 4.8 W. In each panel, we show the evolution map directly obtained from the numerical NLSE simulations and that obtained from the RNN network model. The RNN was trained with 950 normalized NLSE realizations where the soliton number was randomly varied in the range 1 to 7. The r.m.s. error (see Eq. (4) in Methods) computed over 50 test realizations was R = 0.152 and R = 0.077 for the temporal and spectral intensities, respectively.

Extended Data Fig. 3 Inclusion of input pulse chirp.

(a) and (b) shows temporal evolution for input pulse parameter similar to those in Extended Data Fig. 2(a) and (c), but when a linear chirp (parabolic phase) of C = − 5 and C = 5, respectively, was added to the input pulse (see Methods). The left and middle panels show temporal evolution from NLSE simulations and prediction by the RNN, respectively, and the right panel shows the comparison between these two at selected distances. The network was trained with 5900 NLSE realizations where the soliton number and chirp parameter were varied randomly in the range 1 to 7 and -8 to 8, respectively. The r.m.s. error computed over 100 test realizations is R = 0.156.

Extended Data Fig. 4 Generalization of RNN applicability to modelling GNLSE dynamics.

Training data were generated using the normalized form of the GNLSE including Raman effect, self-steepening and third-order dispersion (see Methods). The results are plotted in dimensional units. The network was trained using 11800 normalized GNLSE realizations where the soliton number, normalized third-order dispersion parameter, and pulse duration were respectively randomly varied in the range 2 to 8, 1 to 9 and 30 and 130 fs (FWHM). (a) shows the results for a transform-limited N = 4 input pulse centered at 830 nm with 7.6 kW peak power and 40 fs duration. Corresponding fibre parameters are γ = 0.1 W−1m−1, β2 = − 8 × 10−27 s2m−1 and β3 = 9 × 10−41 s3m−1. (b) shows the results for a transform-limited N = 7 input pulse with 2.9 kW peak power and 120 fs duration. Corresponding fibre parameters are γ = 0.0184 W−1m−1, β2 = − 5.1 × 10−27 s2m−1 and β3 = 4.3 × 10−41 s3m−1. (c) shows the results for a transform-limited N = 4.5 input pulse with 3.0 kW peak power and 60 fs duration. Corresponding fibre parameters are γ = 0.01 W−1m−1, β2 = − 1.7 × 10−27 s2m−1 and β3 = 6.5 × 10−42 s3m−1. In each panel, we show the evolution map directly obtained from the numerical GNLSE simulations and that obtained from the RNN network model. The r.m.s. error computed over 200 test realizations is R = 0.092.

Extended Data Fig. 5 Effect of input noise on RNN predictions.

The left and right panels shows how ± 10 and ± 20 % relative random multiplicative intensity noise added to the examples shown in Extended Data Fig. 2(a) and (c) affect the RNN predictions (see also Methods). The r.m.s. error computed over 50 test realizations was R = 0.200 and R = 0.271 for the ± 10 and ± 20 % cases, respectively (R = 0.152 for noise-free data). Although we do note a residual shift in the point of maximum compression, the dynamics of the higher-order soliton compression are overall well reproduced even under noisy conditions.

Extended Data Fig. 6 Application of RNN to modelling multimode GNLSE dynamics.

RNN modelling of nonlinear propagation of an ultrashort pulse of 150 fs duration (FWHM) and 2.5 MW peak power with 1500 nm center wavelength in a 50 μm radius multimode silica fibre. The network was trained with 950 realizations where the energy distribution between the different modes at the fibre input was varied. (a-d) show the spectral evolution integrated over all the modes for input energy distributions as indicated. In each panel, we show the evolution map directly obtained from the numerical multimode GNLSE simulations and that obtained from the RNN network model. The r.m.s. error computed over 50 test realizations is R = 0.104.

Extended Data Fig. 7 Computation time comparison of GNLSE and RNN modelling.

(a) Simulation time to compute evolution maps of supercontinuum similar to those presented in Fig. 6 using the GNLSE and RNN, as a function of the number of points Np in the temporal (or spectral) grid and number of computed maps as indicated in the parentheses. (b) Simulation time to compute evolution maps of multimode simulations similar to those shown in Extended Data Fig. 6 using the multimode GNLSE and RNN, as a function of the number of realizations for a constant number of modes and grid points.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Salmela, L., Tsipinakis, N., Foi, A. et al. Predicting ultrafast nonlinear dynamics in fibre optics with a recurrent neural network. Nat Mach Intell 3, 344–354 (2021). https://doi.org/10.1038/s42256-021-00297-z

Download citation

Further reading

Search

Quick links