Cavity-Soliton Reservoir Computing

• Cavity-soliton reservoir computing is a photonic architecture that uses nonlinear soliton dynamics and spectral multiplexing to create high-dimensional computational states.
• It exploits the internal degrees of freedom and radiative sidebands of localized optical pulses to enhance memory capacity and nonlinearity without relying solely on external modulators.
• The approach contrasts coherently driven fiber-cavity implementations with gain-assisted micro-combs, achieving improved performance on tasks like XOR, Mackey–Glass prediction, and channel equalization.

A cavity-soliton reservoir computer is a photonic reservoir computing architecture in which the reservoir is furnished by the nonlinear, driven-dissipative dynamics of cavity solitons in an optical cavity. In the direct realization reported to date, the reservoir is the transient response of a temporal cavity soliton and its radiative waves in a coherently driven fiber ring, the input is encoded as phase modulation of the drive laser, and the reservoir state is obtained by frequency-resolved readout of the output spectrum (Arabieh et al., 20 Feb 2026). Related cavity-soliton platforms, notably laser cavity-soliton micro-combs in nested cavities, supply a broader physical context: they provide self-localized pulses, broad combs, multistability, and repetition-rate tunability, all of which suggest a high-dimensional photonic substrate for recurrent analog computation (Bao et al., 2019).

1. Physical basis of cavity-soliton computation

Temporal cavity solitons are localized optical pulses sustained in a resonator by a double balance. In the fiber-cavity implementation used for reservoir computing, dispersion is balanced by Kerr nonlinearity during propagation, while cavity losses are balanced by coherent driving from a CW pump on successive roundtrips. The resulting structure is a driven-dissipative soliton whose peak power, width, and phase are fixed by cavity parameters such as detuning, loss, nonlinearity, dispersion, and pump power. In that system, cavity solitons circulate at the cavity free spectral range while their internal structure remains on the picosecond scale (Arabieh et al., 20 Feb 2026).

A central distinction within the broader cavity-soliton literature is that between conventional Kerr cavity solitons described by the Lugiato–Lefever framework and laser cavity-solitons sustained by gain in a second cavity. Standard micro-comb solitons are temporal cavity solitons localized on a strong CW background set by an externally injected pump, with bright solitons typically carrying 95% of the energy in the CW mode. By contrast, the nested-cavity laser cavity-soliton micro-comb platform is self-sustained by gain in the outer cavity, has a zero background as its low-energy state, does not require an external coherent drive at the comb frequency, and operates as a localized steady state of a two-cavity, multi-mode system (Bao et al., 2019).

This distinction matters computationally. The fiber cavity-soliton reservoir computer uses the perturbative response of a soliton sitting on a CW background, whereas the laser cavity-soliton micro-comb platform shows that background-free localized states, multistability, and repetition-rate tunability can be realized in micro-cavities with average powers one order of magnitude lower than state-of-the-art approaches and with no active feedback (Bao et al., 2019). This suggests a spectrum of cavity-soliton reservoirs ranging from coherently driven passive cavities to gain-assisted nested cavities.

2. Reservoir architecture and state formation

In the fiber-cavity realization, scalar discrete-time input $u(m)$ is encoded into the phase of the CW pump. The phase is held constant for $q$ cavity roundtrips,

$\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$

where $n$ is the roundtrip index, $q$ is the repetition factor, and $M$ is the modulation strength. Each phase step perturbs the soliton away from its stationary state; the soliton then relaxes back through transient amplitude and phase oscillations, with concomitant spectral breathing. The high dimensionality is therefore not produced by a network of separate components but by the internal degrees of freedom of the soliton and, crucially, by the frequency domain (Arabieh et al., 20 Feb 2026).

The reservoir nodes are defined by frequency slices. The output spectrum is partitioned by a programmable spectral filter into $N$ channels, and for each symbol the node value $x_i(m)$ is the average optical power in channel $i$ over the corresponding $q$ roundtrips: $q$0 The readout is linear,

$q$1

with $q$2 obtained by ridge regression. In simulation, filters are randomly distributed over one side of the spectrum to reduce redundancy; experimentally, a Waveshaper selects spectral windows sequentially and a 10 MHz photodiode measures channel power (Arabieh et al., 20 Feb 2026).

The physical implementation used a 60 m SMF-28 fiber ring cavity with $q$3 and $q$4. The cavity comprised passive elements together with 30 cm of Er-doped fiber pumped at 1480 nm, effectively reducing roundtrip loss to $q$5 and yielding finesse $q$6. Cavity solitons were initialized by addressing pulses from a mode-locked laser at 1535 nm injected through a WDM, while detuning stabilization employed a separate low-power CW control beam and PID control of the main laser frequency (Arabieh et al., 20 Feb 2026).

A misconception is that cavity-soliton reservoir computing must be organized as a time-multiplexed single-node delay system. The demonstrated cavity-soliton implementation is instead frequency multiplexed: different wavelengths act as different nodes, all evolving in parallel in the same cavity (Arabieh et al., 20 Feb 2026). Time multiplexing remains relevant as a neighboring design pattern in photonic reservoir computing, but it is not the defining principle here.

3. Governing equations and dynamical structure

The most complete model used for the cavity-soliton reservoir is the Ikeda map. Propagation over one roundtrip obeys the nonlinear Schrödinger equation

$q$7

and the boundary condition at the input coupler is

$q$8

Here $q$9 is the input-coupler coefficient, $\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$0, $\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$1 is the pump power, $\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$2 is the detuning, and $\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$3 carries the input (Arabieh et al., 20 Feb 2026).

When the field varies weakly from roundtrip to roundtrip, the mean-field approximation leads to the Lugiato–Lefever equation,

$\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$4

A reduced variational description then approximates a single cavity soliton by

$\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$5

yielding coupled ODEs for the soliton amplitude $\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$6 and phase $\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$7. The stationary solution satisfies

$\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$8

and implies the detuning bound

$\varphi_n = M\,u(m), \quad (m-1)q+1 \le n \le mq,$9

Linearization around the stationary point produces damped relaxation oscillations with damping time of order $n$0, which sets a memory timescale for the reservoir (Arabieh et al., 20 Feb 2026).

The decisive dynamical distinction is that the mean-field LLE does not support Kelly sidebands, whereas the Ikeda map does. Kelly sidebands arise from periodic perturbations once per roundtrip; in the cavity-soliton reservoir they appear as narrow spectral peaks around the soliton spectrum, and their normalized spectral fluctuations exhibit pronounced maxima at the sideband frequencies. Model comparison shows that the reduced model is essentially one-dimensional, the LLE adds the CW background but still omits discrete sidebands, and only the Ikeda model reproduces the Kelly-wave structure that materially enhances memory and nonlinear task performance (Arabieh et al., 20 Feb 2026).

A related, but distinct, dynamical framework appears in laser cavity-soliton micro-combs. There the micro-ring field $n$1 and outer-cavity super-mode fields $n$2 obey a vector laser mean-field model,

$n$3

$n$4

This system supports self-organized states in which all active super-modes lock to a common group velocity $n$5 and frequency $n$6, providing a multi-mode cavity-soliton substrate with multistability, walking solitons, and tunable repetition rate (Bao et al., 2019).

4. Computational properties and benchmark performance

The cavity-soliton reservoir has been evaluated on linear memory, nonlinear channel equalization, Mackey–Glass prediction, XOR, and Hénon prediction. For linear memory capacity, using the Ikeda model with $n$7, $n$8, $n$9, the maximum $q$0 occurs at $q$1 and $q$2. This is below the node-count upper bound of 50 but sufficiently large to indicate substantial short-term memory. Larger phase modulation eventually destroys the soliton, so the optimal linear-memory regime is relatively close to linear response (Arabieh et al., 20 Feb 2026).

For nonlinear channel equalization at channel SNR $q$3, best numerical accuracy was $q$4 for $q$5, $q$6, $q$7, $q$8, and $q$9. Increasing $M$0 to $M$1 expanded the stable operating region and improved peak accuracy to $M$2 at $M$3 and $M$4. The Mackey–Glass benchmark yielded $M$5 for 6-step-ahead prediction, $M$6 for 10 steps, and $M$7 for 15 steps under the reported conditions. For XOR, the experimental system used $M$8 channels of 120 GHz each at $M$9 and $N$0; higher $N$1 improved short-delay accuracy, whereas smaller $N$2 preserved memory better for longer delays. For Hénon-map one-step prediction, the experiment with $N$3 channels of 60 GHz each at $N$4, $N$5, $N$6, and $N$7 produced $N$8, versus $N$9 for a linear reservoir (Arabieh et al., 20 Feb 2026).

The principal computational lesson is that Kelly waves are functional degrees of freedom rather than parasitic artifacts. In the reported comparisons, reduced and LLE models show poor XOR accuracy and lower memory, whereas the Ikeda model achieves much better performance because the sidebands provide channels with distinct sensitivities and relaxation dynamics. Enhancing Kelly sidebands through localized dispersion changes modifies performance in a task-dependent way, confirming that the reservoir dimension is not only the soliton envelope but also its radiative structure (Arabieh et al., 20 Feb 2026).

A second misconception is that photonic reservoir performance depends only on external nonlinearities in modulation or detection. A coherent all-optical fiber-ring reservoir with distributed Kerr nonlinearity shows that bulk optical nonlinearity increases nonlinear computational capacity beyond what is available from a Mach–Zehnder input nonlinearity or photodiode readout alone. In that system, adding Kerr nonlinearity produced significant degree-3 and degree-4 memory capacities and reduced Santa Fe prediction NMSE to below $x_i(m)$0 in cases where a linear reservoir remained markedly worse (Pauwels et al., 2019). This does not make the cavity-soliton substrate redundant; rather, it clarifies why an intrinsically nonlinear cavity can be expected to outperform architectures whose cavity is linear and whose nonlinearity is confined to the I/O stack.

5. Position within photonic reservoir computing

Cavity-soliton reservoir computing sits within a larger family of cavity-based photonic reservoirs. A coherently driven passive fiber cavity can already function as a high-performance reservoir even when the cavity itself is linear, with the nonlinearity supplied by quadratic photodetection. In that passive-cavity architecture, the field obeys

$x_i(m)$1

the virtual-node states are sampled by time multiplexing, and the readout uses

$x_i(m)$2

With $x_i(m)$3, the reported experimental memory capacities were $x_i(m)$4, $x_i(m)$5, $x_i(m)$6, and $x_i(m)$7, while the same platform attained $x_i(m)$8 word error rate on noiseless isolated spoken digit recognition with $x_i(m)$9 and $i$0 on NARMA10 with $i$1 (Vinckier et al., 2015). The cavity-soliton reservoir differs in that the reservoir nonlinearity is intrinsic to the cavity dynamics itself.

Another neighboring line is memory-reconfigurable time-delay photonic reservoir computing. In a resonant-cavity architecture based on an asymmetric Mach–Zehnder interferometer integrated in the cavity, the memory capacity can be tuned without an optical attenuator block, and the approach can be leveraged to find the optimal value for the specific components of the total memory capacity metric. That design was demonstrated on the temporal bitwise XOR task and concluded that memory-capacity reconfiguration allows optimal performance for memory-specific tasks (Abdalla et al., 2022). This is not a cavity-soliton device, but it addresses the same system-level question of how memory is physically tuned in a cavity reservoir.

The micro-comb route extends the cavity-soliton idea toward integrated multi-mode sources. In the nested-cavity laser cavity-soliton platform, the micro-ring has $i$2, $i$3, anomalous $i$4, and $i$5; the outer cavity has $i$6, a 10 nm bandpass filter with $i$7, and anomalous $i$8. The system demonstrated $i$9 comb bandwidth, injected average power at $q$0 and $q$1 of the LLE threshold in fitted cases, and repetition-rate adjustments of $q$2 and $q$3 by changing only the outer-cavity length. The paper explicitly identified amplitude, phase, timing or velocity, center frequency, and multi-soliton states as relevant degrees of freedom, and it discussed spectral, temporal, and hybrid temporal–spectral reservoir readout strategies as plausible computing modes (Bao et al., 2019).

Accordingly, “cavity-soliton reservoir computer” should not be conflated with all cavity-based reservoir computers, nor restricted to one implementation. The established instances span passive coherent cavities, distributed-Kerr fiber rings, memory-reconfigurable resonant cavities, frequency-multiplexed driven-soliton fiber cavities, and nested-cavity laser cavity-soliton micro-combs (Vinckier et al., 2015).

6. Limitations, open questions, and future directions

Current cavity-soliton reservoir experiments remain constrained by platform-specific bottlenecks. In the demonstrated fiber system, symbol rates were in the kHz range because of AWG and locking limitations together with relatively large $q$4, even though the cavity FSR is in the MHz range. Node-wise SNR rises with modulation depth $q$5, but spectral channels close to the pump and at spectral tails are disadvantaged: near the pump the relative spectral variation is small, while at the tails the optical power is low. Numerical simulations that reproduce the experiments use node-wise SNR of approximately $q$6 (Arabieh et al., 20 Feb 2026).

There are also questions of physical generality. Kelly sidebands are naturally strong in the strongly coupled fiber cavity used for the first demonstration, but they are weaker in microresonators because coupling is small and the mean-field approximation is more accurate. The current evidence therefore supports a qualified conclusion: cavity solitons are already sufficient to produce a useful nonlinear reservoir, but in the reported fiber implementation the best performance depends materially on radiative structure beyond the mean-field soliton core. It would be inaccurate to reduce the substrate to a bare LLE soliton alone (Arabieh et al., 20 Feb 2026).

Several development paths are explicit in the literature. For the cavity-soliton reservoir itself, proposed directions include analog output layers, deliberate enhancement of Kelly waves by tailored dispersion or cavity design, microresonator implementations, improved comb efficiency and flatness, multi-soliton reservoirs, and combined time-and-frequency multiplexing (Arabieh et al., 20 Feb 2026). For laser cavity-soliton micro-combs, the reported future directions include genetic algorithms for self-starting and adaptive control and the exploitation of dispersive wave coupling for broader spectra (Bao et al., 2019). A plausible implication is that future cavity-soliton reservoirs will not be defined by a single canonical cavity equation or readout geometry, but by a shared operating principle: localized driven-dissipative structures furnish recurrent memory and nonlinear transformation, while spectral, temporal, or hybrid multiplexing supplies the observable state space.

A final misconception is that cavity-soliton reservoirs are merely optical analogues of conventional delay reservoirs with a different nonlinearity block. The available evidence points to a stronger claim: the soliton, its relaxation spectrum, and in some platforms its radiative sidebands are themselves the computational substrate. That is why comparisons among reduced models, mean-field LLE models, full roundtrip maps, passive coherent cavities, and gain-assisted nested cavities are not incidental. They identify which physical degrees of freedom actually supply memory, nonlinearity, and separability, and therefore determine what a cavity-soliton reservoir computer is in practice (Arabieh et al., 20 Feb 2026).