Source-linked AI summary
Deep learning for dynamic modeling and coded information storage of vector-soliton pulsations in mode-locked fiber lasers
Zhi-Zeng Si, Da-Lei Wang, Bo-Wei Zhu, Zhen-Tao Ju, Xue-Peng Wang, Wei Liu, Boris A. Malomed, Yue-Yue Wang, Chao-Qing Dai
TL;DR
Modeling nonstationary soliton dynamics in mode-locked fiber lasers is computationally demanding and sensitive to parameter variations. The paper uses a TP-Bi_LSTM RNN to predict vector-soliton pulsations from operational signal data, reports experimental verification and better results than direct simulations for specified regimes, and demonstrates coded-information storage using predicted signals.
Problem
Numerical modeling of nonstationary mode-locked fiber-laser solitons requires accurate parameter mapping, is sensitive to small input variations, and can be inefficient.
Source: §1 · p. 2Method
A TP-Bi_LSTM RNN with attention predicts subsequent vector-soliton evolution from a small amount of initial data while the laser is operational.
Source: §1 · pp. 2–3Results
RNN prediction RMSE was lower than CNLSE simulations, and predictions for period-21 and period-3/period-43 VSPs were better than direct simulations and experimentally verified.
Source: §1 · p. 3, §3 · p. 14Takeaways & Limitations
Using predicted rather than experimental pulse signals for coded-information storage avoids the losses and low stability associated with actual-pulse transmission.
Source: §3 · p. 15Numerical simulations remain necessary to generate training samples, including a 10^-6–10^-8 relative perturbation to the gain-fiber group-velocity-dispersion coefficient that is difficult to implement experimentally.
Source: §2.1 · p. 3
Abstract
from arXiv · showhide
Soliton pulsations are ubiquitous feature of non-stationary soliton dynamics in mode-locked lasers and many other physical systems. To overcome difficulties related to huge amount of necessary computations and low efficiency of traditional numerical methods in modeling the evolution of non-stationary solitons, we propose a two-parallel bidirectional long short-term memory recurrent neural network, with the main objective to predict dynamics of vector-soliton pulsations in various complex states, whose real-time dynamics is verified by experiments. Besides, the scheme of coded information storage based on the TP-Bi_LSTM RNN, instead of actual pulse signals, is realized too. The findings offer new applications of deep learning to ultrafast optics and information storage.
1. Introduction
The introduction motivates modeling complex vector-soliton pulsations in mode-locked fiber lasers and identifies limits in existing numerical and neural-network approaches. It presents TP-Bi_LSTM prediction of operational laser signals and coded-information storage using predicted pulses.
1. Introduction
- VSPs are pulsating two-component solitons with periodic energy exchange, and their complex dynamics is relevant to chaos, spectral analysis, and fiber-optic communications.The introduction describes VSP transitions toward dynamical chaos as involving sudden period changes caused by higher-order dispersion and nonlinearity.Source: §1 · pp. 1–2
- SSFM modeling requires accurate cavity-to-NLS parameter mapping, while nonstationary outputs are sensitive to small input changes and simulations are inefficient.The introduction also notes that PINN is unsuitable for predicting long-distance MLFL optical-signal evolution because it is strongly correlated with the modeled equations.Source: §1 · p. 2
- Although LSTM had been used to model optical supercontinuum, RNN prediction had not been applied to nonstationary MLFL soliton dynamics.Prior MLFL feedforward and residual convolutional networks were not consistently compared with experimental findings.Source: §1 · p. 2
- TP-Bi_LSTM RNN with attention predicts subsequent nonstationary vector-soliton evolution from a small amount of initial data once the laser is operational.The predictions are reported as consistent with transient experimental information and as reducing calculation and time costs.Source: §1 · pp. 2–3
- Predictions of stable and unstable VSP regimes agree with transient experimental information obtained using TS-DFT.The introduction describes the predictions as highly accurate and as reducing calculation and time costs.Source: §1 · p. 3
- Coded-information storage is realized using RNN-predicted VSP pulse signals instead of actual pulse signals, with experimental verification.The coding system is based on different VSPs and uses predicted pulse signals verified by experiment.Source: §1 · p. 3
2. Results
The TP-Bi_LSTM RNN predicts periodic, bi-periodic, and unstable vector-soliton pulsations from initial pulse data, with predictions compared against simulations and experiment. Predicted pulse sequences are also used to demonstrate coded information storage.
2.1 Procedure
- The procedure trains the supervised RNN on CNLSE-generated datasets, tests predictions against experiments, and then uses verified predicted pulse signals for coding.Training uses 500 samples with 1500 RTs each, split into 450 training and 50 testing samples; prediction is based on a small number of initial pulses.Source: §2.1 · pp. 3–4
2.2.1 The dynamical prediction and experimental verification of the periodic VSP
- For period-21 pulsations, RNN predictions match experiment and outperform CNLSE simulations: component U and V deviations are 36% and 13% lower, respectively.Predicted pulse widths are 1.54×490 fs, close to the experimental 1.54×487 fs; predicted transient spectra also closely agree with experiment.Source: §2.2.1 · pp. 7–8, Figure 3 · p. 7
2.2.2 The dynamical prediction and experiment verification of the bi-periodic VSP
- For the period-3/43 bi-periodic VSP, RNN predictions agree with experiment and have an RMSE of 0.045, versus 0.055 for CNLSE simulations.The prediction's deviation is 18% less than the simulations'; the predicted and experimental pulse widths are both 1.54×732 fs.Source: §2.2.2 · pp. 9–10, Figure 4 · p. 9
- For period-3/43 pulsations, prediction error remains within 0.2 for the first 3900 RTs of a 5000-RT forecast, then reaches 0.31 at RT 5000.The mean absolute error is 0.059 over the first 3900 RTs and 0.082 over all 5000 RTs; the training loss converges to approximately 10^-7.Source: Figure 5 · p. 9, §2.2.2 · p. 10
- Predictions for period-3/18 and period-2/18 bi-periodic VSPs agree with experiment; their RMSEs are 0.155 and 0.085, respectively, across 127 RTs.The period-2/18 error and 80% of the period-3/18 error remain within 0.2.Source: Figure 6 · p. 10, §2.2.2 · p. 11
2.2.3 Unstable state of VSP
- For unstable period-9 pulsations, the oscillation period varies from 11 to 9 RTs during formation; after stabilization, RT error falls below 0.1.The authors report that selecting a suitable training set and increasing training iterations improve prediction performance in unstable states.Source: §2.2.3 · pp. 11–12, Figure 7 · p. 11
- The trained RNN produces prediction data in about 95 seconds, compared with CNLSE simulations taking dozens of hours; simulations keep the CPU at full load.The authors also report reduced CPU dependence for the RNN.Source: §2.2.3 · pp. 12–13, Table 1 · p. 12
3. Conclusion
TP-Bi_LSTM predictions of complex vector-soliton pulsations were verified against real-time experimental dynamics and outperformed direct simulations for two reported states. The predicted signals also formed the basis of a real-time pulse-coding storage system.
3. Conclusion
- 36% and 18% less deviation than direct simulations for period-21 VSP and bi-periodic period-3/period-43 VSP, respectively.The RNN prediction had lower RMSE than CNLSE simulations when compared with experiment.Source: §3 · p. 14
- Real-time transient experiments verified predicted periodic and multi-period VSPs; suitable training data and more iterations improved accuracy for the unstable period-9 state.Source: §3 · p. 14
- Using predicted rather than experimental pulse signals, the real-time coding system avoids losses and low stability associated with actual optical-pulse transmission.The system was designed on the basis of the predicted RNN signal.Source: §3 · pp. 14–15
- Future work proposes optimizing laser operating parameters to realize specific soliton states and improve laser output efficiency.Source: §3 · p. 15