Neural network predicts Kerr resonator solitons 20 times faster than simulations

Researchers have shown that a recurrent neural network can reproduce the intricate dance of light inside a Kerr resonator roughly twenty times faster than the traditional numerical simulations that have long dominated the field. The study, published as an open access research article in Frontiers of Optoelectronics on 25 September 2025, tackles a well-known bottleneck: modeling the nonlinear dynamics of these ring-shaped optical cavities usually means solving the Lugiato–Lefever equation with the split-step Fourier method, a computationally intensive and time-consuming procedure. By training a gated recurrent unit network on simulation data and feeding prior information back into the model, the team achieved efficient and accurate prediction of soliton dynamics, with graphics processing unit acceleration delivering the twenty-fold gain in computational efficiency.

Kerr resonators are among the most popular platforms for producing optical frequency combs and temporal cavity solitons. Temporal cavity solitons are self-sustained, localized light pulses that circulate indefinitely within an externally driven Kerr resonator, maintaining their shape through a composite balance between cavity dispersion and Kerr nonlinearity, together with parametric gain and loss. Because they can be integrated on chip, these pulses hold significant potential for applications in optical memories and all-optical systems, and they have attracted considerable attention since their discovery. Describing their dynamics typically relies on the Lugiato–Lefever equation, a simplified description of nonlinear and dispersive effects that captures the behavior of nonlinear ring resonators, but solving it with the split-step Fourier method demands extensive computational work that hinders research on temporal cavity solitons.

Applying machine learning to this problem is not straightforward. The nonlinear dynamics in ring cavities involve processes at multiple different time scales, which increases the complexity of analysis, and complex phenomena such as chaotic states exist within the cavities. The strong randomness of chaos makes it difficult for artificial intelligence to learn and predict, and the authors note that applying machine learning to the study of dynamics in nonlinear ring cavities has remained an unresolved issue. Their approach uses a recurrent neural network to model the nonlinear dynamics within a Kerr fiber ring resonator, employing a prior information method that lets the model make accurate judgments under different cavity parameters.

The first phase of the work addressed a resonator driven by a single pump. The simulated system was a fiber Kerr resonator built from 100 meters of single-mode fiber closed by a 90/10 coupler. A pulse-pattern generator produced high-precision Gaussian pulses that modulated a continuous-wave light source through an amplitude modulator, while an optical isolator suppressed stimulated Brillouin scattering and a 99/1 tap coupler allowed monitoring of the intracavity dynamics. The optical field evolution was governed by the dimensionless mean-field Lugiato–Lefever equation, in which successive terms represent cavity losses, Kerr nonlinearity, cavity phase detuning, group velocity dispersion, and rapid time-varying coherent driving, with the cavity driven by Gaussian optical pulses under anomalous dispersion.

Three adjustable parameters defined the training data: the pump pulse width, ranging from 15 to 21 in steps of 1; the amplitude of the Gaussian driving field, from 1.5 to 2.1 in steps of 0.1; and a scanned detuning from 0.004 to 0.016 in steps of 0.002 per round trip. Beyond the specified pulse width range the cavity becomes unstable and enters a chaotic state, so the scanning speed was kept within this range to save memory and analyze the soliton formation process in detail. A maximum of 1500 round trips ensured a stable state. In total 343 data sets were generated, of which 200 were randomly selected for training; from the remaining 143, seventy were chosen as the test set.

The model architecture passes segmented two-dimensional arrays through a dense layer, then uses broadcasting to expand three-column parameter vectors to 512 columns, constructing a recurrent network with the sampling points. The data then flows into gated recurrent unit layers, with two GRU layers and a sequence length of twelve: the twelfth time sequence is predicted from the previous eleven, and the predicted sequence becomes the next input in a sliding window prediction mode. Accuracy was evaluated with the root mean square error between predicted field intensity and the Lugiato–Lefever result, computed per round trip and averaged across all round trips and samples. Among the recurrent architectures compared, the gated recurrent unit demonstrated superior performance for this task.

The predictions tracked the physics closely. The model successfully reproduced the formation of a single soliton and, for the Gaussian pumping scheme, the alternative state of dual solitons, matching the evolution obtained from the Lugiato–Lefever equation in both cases. When the pump pulse width and driving amplitude increased, Turing rings emerged during the initial stage of intracavity evolution, transitioning into a chaotic state and eventually into a single soliton. The model accurately predicted the formation of the Turing rings and their transition into chaos; because of the random nature of the chaotic regime it could not fully predict the dynamics within it, but the evolution trend matched well. The authors attribute this pattern to the fact that single solitons, double solitons, and Turing rings possess stable physical structures while chaotic states are strongly random, so neural networks more readily capture the ordered regimes.

The second phase extended the method to bichromatic driving, targeting parametrically driven cavity solitons. Unlike conventional cavity solitons, which are driven by a monochromatic continuous-wave field and draw energy from the background through four-wave mixing, parametrically driven cavity solitons rely on a bichromatic driving field and form through phase-sensitive amplification based on non-degenerate four-wave mixing. Their spectral center lies between the two driving frequencies, they exhibit two opposite phase-stable states, and these properties endow them with enhanced noise resistance and superior phase controllability. The simulated resonator had a cavity length of approximately 8.3 millimeters, a nonlinear coefficient of 1.2 W−1·km−1, a finesse of 5000, and both second- and fourth-order dispersion, with two continuous-wave driving fields of 200 milliwatts each. Pump power was varied from 0.10 to 0.20 watts and detuning from 0.5 to 1.6 radians, generating 132 data sets split into 100 training, 16 validation, and 16 test sets.

For this bichromatic case the team combined a convolutional neural network, which extracts representative features from the detuning and pump power inputs, with the gated recurrent unit that handles temporal sequences and long-term dependencies. A sequence length of eight and two GRU layers yielded favorable results. When the pump power was 0.2 watts and the detuning 1.2, the intracavity field satisfied the three conditions for parametrically driven cavity soliton formation, and the model reproduced the stable soliton state, matching the Lugiato–Lefever comparison at the 6500th round trip. In benchmark comparisons, the CNN-GRU network achieved a significantly lower root mean square error than a plain GRU with only a slight increase in parameters, and required just 3.49 seconds on a graphics processor, nearly twenty times faster than the traditional model; even without GPU acceleration it remained almost six times faster, although a plain GRU was then more efficient because of the computational demands of the convolutional layers.

The broader significance, the authors argue, is that artificial intelligence can model nonlinear optical dynamics in Kerr resonators and pave the way for designing optical frequency combs and generating ultrafast pulses. Notably, the speed advantage is structural rather than incidental: GPU acceleration is limited for the split-step Fourier method because its computation has strong temporal step dependency, with each step relying on the previous one, which prevents full use of parallel processing suited to large-scale independent tasks. The recurrent approach, by contrast, benefits directly from parallel hardware. The team hopes the method will become a useful modeling tool for Kerr nonlinear ring resonators, facilitating control of temporal cavity solitons and opening possibilities for future all-optical systems and optical memory.

Subject of Research: Using recurrent neural networks to rapidly predict nonlinear soliton dynamics in Kerr resonators

Article Title: Rapid prediction of complex nonlinear dynamics in Kerr resonators using the recurrent neural network

Article References: Huang, T., Chen, L., Lu, M., Pan, J., Xu, C., Wang, P., & Shum, P. P. (2025). Rapid prediction of complex nonlinear dynamics in Kerr resonators using the recurrent neural network. Frontiers of Optoelectronics, 18(1), Article 19. https://doi.org/10.1007/s12200-025-00164-4

Image Credits: AI Generated

DOI: 10.1007/s12200-025-00164-4

Keywords: Kerr resonator, recurrent neural network, gated recurrent unit, temporal cavity solitons, Lugiato–Lefever equation, optical frequency comb, GPU acceleration, bichromatic driving, parametrically driven cavity solitons, Turing rings, machine learning, nonlinear optics