betaFeedback

WorksPhysics and Astronomy

Polarization Lifetime as a Design Axis for Task-Adaptive Photonic Reservoir Computing

Abstract

Photonic time-delayed reservoir computing (TDRC) based on conventional Class B lasers is limited by fragile external cavity mode (ECM) locking dynamics. Here, we show that engineering the polarization lifetime (T2_2) with the carrier lifetime (TT) reshapes the topology and stability of TDRC. Comparing four laser classes (Class A, Class B, Class CA_A, and Class CB_B), a finite polarization lifetime acts as an intrinsic dynamic filter, dissolving discrete ECM chaos grids into continuous, phase-tolerant operational basins. Information processing capacity (IPC) decomposition further exposes distinct computational profiles: matched Class CA_A achieves the highest nonlinear capacity (16.28) and the highest total capacity (19.29), whereas Class A exhibits the highest linear capacity (4.70). A task-dependent reversal governs hardware design: Class A and Class CA_A achieve comparable, and the lowest, NRMSE on short-horizon Mackey–Glass prediction (0.0103 and 0.0105, within one standard deviation), while the time-scale-matched Class CA_A dominates the higher-dimensional Lorenz (0.0506) and the experimental Santa Fe (0.4251) tasks. Linear-IPC ranking matches Mackey-Glass accuracy; nonlinear-IPC ranking matches the Lorenz exactly and Santa Fe within the mask-to-mask uncertainty. Under realistic quantum noise, resilience is dictated mainly by TT, not T2_2: Class A retains the lowest NRMSE (0.47) even at spontaneous emission factor ββ = 10−110^{-1}, whereas the long-carrier Class B degrades (0.98). A residual T2_2-mediated benefit of 20%\% persists in the slow-carrier Class CB_B (0.78 vs. 0.98). Time-scale engineering thus offers a nuanced design paradigm for optical computing, where optimal hardware is set by the task's time-scale hierarchy rather than a universal figure of merit.

PACS numbers:

Introduction

The exponential growth of artificial intelligence, particularly in time-series forecasting and real-time control, has created an urgent demand for neuromorphic hardware capable of processing temporal data streams with ultrahigh throughput and minimal energy consumption [1, 2]. Photonic reservoir computing (RC) has emerged as a promising paradigm to meet these demands [3]. By harnessing the transient, high-bandwidth dynamics of optical systems to project input signals into a high-dimensional state space, photonic RC circumvents the computationally expensive iterative training of recurrent neural network weights [4, 5]. In the widely adopted time-delayed RC (TDRC) architecture, the spatial complexity of a conventional network is elegantly replaced by the temporal multiplexing of a single physical nonlinear node subject to delayed optical feedback [6, 7, 8]. Consequently, the computational capacity, memory depth, and noise resilience of the entire system are dictated entirely by the intrinsic physical dynamics of this single node.

To date, the vast majority of photonic TDRC implementations have relied on semiconductor lasers as the nonlinear node [4, 9, 10, 11, 12, 13, 14]. Experimentally, photonic TDRC has achieved high-throughput processing at multi-gigabit rates using transients in delayed feedback systems [13]. Theoretically, the rich dynamics of semiconductor lasers subject to optical feedback have been extensively modeled using delay differential equations to design optimal reservoir states [14]. The dynamics of such lasers are universally modeled by Class B rate equations, a simplified mathematical framework that assumes the macroscopic medium polarization decays infinitely fast compared to the cavity photon lifetime and the carrier inversion lifetime TT [15, 16]. By adiabatically eliminating the polarization dynamics, Class B models reduce the system to a two-dimensional phase space dominated by relaxation oscillations [17, 18]. While these lasers have been successfully applied to benchmark tasks, their pronounced, carrier-induced relaxation oscillations (which are typically underdamped in conventional semiconductor quantum-well lasers) impose fundamental constraints on the accessible dynamical responses. Most critically, the high-dimensional dynamics close to the edge of stability in Class B lasers are accessible only within narrow windows of a highly structured parameter space, where stable and chaotic dynamics alternate frequently owing to the complex locking of external cavity modes (ECMs). This operational intricacy implies that even minute thermal drifts or fabrication imperfections can push the system out of its optimal computational window, leading to a catastrophic collapse in performance. These fragilities stem from an inherent mismatch between the ultra-fast polarization and slow carrier relaxation timescales.

A promising route is to systematically engineer the laser’s intrinsic timescales, deliberately retaining a finite polarization lifetime (T2T_2) that acts as a dynamical filter and a topological smoother. The adiabatic elimination of the polarization lifetime in Class B models completely discards a rich dynamical resource that could fundamentally alter this picture. This omission has been shown to be particularly significant in the context of the modulation response of nanolasers, where standard rate-equation approaches miss essential dynamical features [19]. When the polarization lifetime (T2T_2) becomes comparable to or longer than the cavity photon lifetime, the laser enters the coherent Class C regime, where polarization evolves as an independent dynamic variable [17]. Moreover, by decoupling T2T_2 from the carrier lifetime TT, one obtains a distinct Class CB_B regime, where a long T2T_2 acts as a natural low-pass filter, imposing a smoothing effect on the complex dynamics of conventional Class B lasers. The role of such slow polarization dynamics as an intrinsic spectral filter has been previously discussed in the context of delay-coupled laser systems [20]. This raises a pressing and previously unanswered question: can the deliberate reintroduction of macroscopic polarization and its interplay with carrier lifetimes fundamentally transform the computational landscape of photonic TDRC? Specifically, can a finite polarization lifetime dissolve the rigid phase-matching constraints that render Class B reservoirs fragile, endow the system with intrinsic adaptability to diverse task timescales, and generate high-dimensional nonlinear mappings that are genuinely robust to spontaneous emission noise?

In this work, we answer this by conducting a systematic comparative study of four representative laser classes: Class A (fast T2T_2, fast TT), Class B (fast T2T_2, slow TT), Class CA_A (finite T2T_2, finite TT), and Class CB_B (finite T2T_2, slow TT). Here, it is necessary to mention that these labels do not strictly follow the classical Arecchi classification. As illustrated in Figure 1, our central thesis is that the deliberate retention of finite T2T_2 acts as a geometric timescale toolkit for shaping the reservoir’s state space. We perform comprehensive sweeps across feedback strength κ\kappa, detuning δω\delta\omega, and phase ϕ\phi to map the operational phase diagrams for each class, revealing that Class CB_B lasers benefit from polarization-induced filtering that significantly broadens their stable regions compared to the fragile locking grids of Class B, namely the strictly phase-matched stable continuous-

Architecture of the photonic TDRC system

Figure 1. Architecture of the photonic TDRC system. The physical hardware core is a laser subject to delayed optical feedback. The four operational regimes - Class A, Class B, Class CA_{\mathrm{A}}, and Class CB_{\mathrm{B}} - are defined by their intrinsic polarization (T2T_2) and carrier (TT) lifetimes. The output signal is routed to a computational path for virtual node generation and readout, and to an experimental path for coherence diagnostics via the second-order correlation function g(2)(0)g^{(2)}(0).

wave (CW) operation locked onto external cavity modes (ECMs) [20]. Harnessing these findings, we lock each class to its optimal parameters, validate task-specific performance on Mackey-Glass (MG), and decompose the computational mechanisms via the rigorous Information Processing Capacity (IPC) framework. We also quantify noise resilience by scanning the spontaneous emission factor β\beta, demonstrating that while Class B’s long carrier lifetime makes it vulnerable to noise accumulation, the natural filtering of Class CB_B provides intrinsic robustness against quantum fluctuations. Collectively, our results establish that engineering laser timescales, particularly the retention of finite T2T_2, is not merely a theoretical exercise, but a powerful practical lever for designing high-performance, noise-resilient, and task-adaptive photonic TDRC hardware.

Theoretical Framework and Laser Dynamics

The foundation of modern photonic RC rests on the rich and complex dynamics of laser systems, which serve as the physical reservoir for information processing. As illustrated in the overview diagram of Figure 1, semiclassical laser models can be broadly categorized into four fundamental operational regimes: Class A, Class B, Class CA_A, and Class CB_B [21]. In addition, quantum dot (QD) lasers often exhibit dynamics that evolve in between Class A and Class B, characterized by relatively faster polarization relaxation than that in conventional quantum-well devices while retaining significant carrier-induced memory effects [21]. The crucial distinction among these configurations lies in the deliberate retention or elimination of the macroscopic polarization lifetime T2T_2, which fundamentally shapes the interplay between the field and the gain medium. A detailed understanding of these dynamics is essential for designing optimal reservoir computing architectures.

The Unified Maxwell-Bloch Framework

The dynamics of the laser-based reservoir are governed by the Maxwell-Bloch delay-differential equations (DDEs), which describe the self-consistent interaction of the complex optical field E(t)E(t), macroscopic polarization P(t)P(t), and carrier inversion N(t)N(t). The equations are expressed as:

dEdt=cP−12E+κE(t−τ)e−iϕ+β2 ξ(t)(1)\frac{\mathrm{d}E}{\mathrm{d}t}=cP-\frac{1}{2}E+\kappa E(t-\tau)e^{-\mathrm{i}\phi}+\sqrt{\frac{\beta}{2}}\,\xi(t) \tag*{(1)}
dPdt=1T2[−(1+iδω)P+EN](2)\frac{\mathrm{d}P}{\mathrm{d}t}=\frac{1}{T_2}\left[-(1+\mathrm{i}\delta\omega)P+EN\right] \tag*{(2)}
dNdt=1T[Ppump(t)−N−2cRe⁡(E∗P)](3)\frac{\mathrm{d}N}{\mathrm{d}t}=\frac{1}{T}\left[P_{\mathrm{pump}}(t)-N-2c\operatorname{Re}(E^*P)\right] \tag*{(3)}

Here EE denotes the complex optical field. The coefficient cc represents the normalized cavity decay rate given by

c=(2δωT2+2)2+1(4)c = \left(\frac{2\delta\omega}{T_{2}}+2\right)^{2}+1 \tag*{(4)}

which ensures the first laser threshold at Ppump=0.5P_{\mathrm{pump}} = 0.5 (δω\delta\omega is the frequency detuning between the cavity mode and the gain resonance). The term KE(t−τ)e−iϕK E(t-\tau)e^{-i\phi} is the time-delayed feedback, which is essential for creating memory and coupling between virtual nodes. PP is the macroscopic polarization with lifetime T2T_{2}. This parameter fundamentally differentiates the laser classes: in Class A, T2T_{2} is nearly instantaneous (T2≪TT_{2} \ll T), effectively eliminating polarization dynamics, whereas in Class CA_{\mathrm{A}} and Class CB_{\mathrm{B}}, a finite T2T_{2} (comparable to the cavity photon lifetime, such as those governed by spin-relaxation rates in semiconductor VCSELs or the prolonged photon lifetimes in high-β\beta microcavities [22]) preserves polarization as an independent dynamical variable. NN is the carrier inversion with lifetime TT, which is driven by the modulated pump Ppump(t)=Pbase(1+γu(t)MN)P_{\mathrm{pump}}(t) = P_{\mathrm{base}}(1+\gamma u(t)M_{N}), where u(t)u(t) is the input signal, MNM_{N} is the periodic mask drawn from a uniform distribution on [−1,1][-1,1], and γ\gamma controls the modulation depth. ξE(t)\xi_{E}(t) is a delta-correlated complex Gaussian noise term with strength β\beta [10], describing spontaneous emission, β=βnoise\beta= \beta_{\mathrm{noise}}.

Time-Delayed Reservoir Computing (TDRC) Implementation

Building on the distinct physical dynamics of the four laser classes, the TDRC architecture fundamentally shifts the traditional neural network paradigm: instead of a spatially distributed network of physical nodes, a single nonlinear node (the laser) with delayed feedback is employed. The implementation is schematically shown in Figure 1. This time-multiplexing scheme was first introduced by Appeltant et al. [4] and has been intensively studied since then [7, 12, 14].

The architecture comprises three primary functional layers: input, reservoir, and output. In the input layer, the temporal input signal u(t)u(t) is sampled and held, then multiplied by a time-multiplexed mask MNM_{N} composed of NN independent random values drawn from a uniform distribution on [−1,1][-1,1], each held constant for the duration of the virtual node spacing θ\theta. In this work, the mask values are drawn from a uniform distribution over [−1,1][-1,1], and the input signal is normalized to the same symmetric range, ensuring that no artificial bias is introduced into the pump modulation. This masked signal directly modulates the laser’s driving current (pump), thereby injecting the input information into the nonlinear physical reservoir. Importantly, this time-multiplexed masking breaks the symmetry of the input signal, discretizes time, and maps the one-dimensional sequence into a high-dimensional virtual-node state space.

The masked input is injected into the laser, perturbing its carrier density and driving the optical field into a transient, nonlinear state. The “reservoir” is constructed using a delayed optical feedback loop characterized by three control parameters: delay τ\tau, feedback strength KK, and feedback phase ϕ\phi. It is worth noting that the present TDRC architecture operates in a non-standard regime in which the feedback delay τ\tau is not necessarily equal to the mask period NvθN_{v}\theta. In the standard TDRC protocol, one sets τ=Nvθ\tau= N_{v}\theta to ensure cyclic coupling between virtual nodes. Here, however, the optimal node spacing θopt\theta_{\mathrm{opt}} is found to scale with the intrinsic relaxation oscillation period τro\tau_{\mathrm{ro}}, which differs by more than an order of magnitude across the four laser classes. Consequently, for a fixed τ=50\tau= 50 and Nv=20N_{v} = 20, the matching condition would require θ=2.5\theta= 2.5, whereas the optimal values range from θ≈1.3\theta\approx1.3 for Class A to θ≈18.6\theta\approx18.6 for Class CB_{\mathrm{B}}. We therefore deliberately allow τ≠Nvθ\tau\ne N_{v}\theta, which generalizes the TDRC concept: the delayed feedback still provides fading memory, but the coupling between virtual nodes becomes local when τ<Nvθ\tau< N_{v}\theta and long-range when τ>Nvθ\tau> N_{v}\theta. This non-standard operation does not invalidate the reservoir computing paradigm; rather, it exploits the intrinsic timescale hierarchy of each laser class. Because the ratio τ/(Nvθ)\tau/(N_{v}\theta) is no longer fixed, tuning θ\theta controls the coupling range between virtual nodes and avoids the degeneracy that arises when the delay is perfectly synchronized with the clock cycle [23]: the delayed field represents the laser’s past state and couples the current virtual node to previous nodes.

The extraction of processed information occurs via the computational path. The laser’s optical output intensity is sampled at intervals θ\theta, producing a high-dimensional state vector of NvN_{v} virtual nodes per delay period. Only the readout weights WoutW_{\mathrm{out}} are trained using ridge regression to minimize the prediction error. This makes training exceptionally fast compared to standard recurrent neural networks.

Benchmark Tasks: Mackey-Glass, Lorenz, and Santa Fe

To quantitatively evaluate the computational capacity, nonlinearity, and memory of the proposed TDRC framework across the four laser classes, we employ three benchmark tasks: the Mackey-Glass (MG) chaotic time series, the three-dimensional Lorenz system, and the Santa Fe laser intensity dataset. For reproducibility, the precise parameters and preprocessing steps of each task are specified below.

Mackey-Glass Chaotic System

The MG series exhibits chaotic dynamics in the chosen parameter regime and is a well-known non-Markovian signal, fundamentally driven by a time-delayed nonlinear differential equation [4]:

dxdt=βMGx(t−τMG)1+[x(t−τMG)]n−γMGx(t)(5)\frac{dx}{dt}=\beta_{\mathrm{MG}}\frac{x(t-\tau_{\mathrm{MG}})}{1+[x(t-\tau_{\mathrm{MG}})]^n}-\gamma_{\mathrm{MG}}x(t) \tag*{(5)}

with the standard chaotic parameters βMG=0.2\beta_{\mathrm{MG}}=0.2, γMG=0.1\gamma_{\mathrm{MG}}=0.1, τMG=17\tau_{\mathrm{MG}}=17, and n=10n=10. In this work, we integrate the MG system with a step size of Δtint=0.001\Delta t_{\mathrm{int}=0.001}, sample it every 0.1 time units, and normalize the input sequence to the range [−1,1][-1,1]. The reservoir is tasked with predicting the state one step ahead, corresponding to a physical time horizon of 0.1 (one sampling interval). This choice ensures that the task demands substantial fading memory capabilities [4, 14]. We emphasize that these parameters differ from those used in earlier literature; therefore, the absolute NRMSE values reported here are not directly comparable to previous benchmark results. Our focus is on the relative performance across the four laser classes under identical conditions.

This task is widely adopted for benchmarking physical reservoirs, as accurate prediction of future values requires the system to maintain fading memory of previous states. This continuous stream is then discretized into time-multiplexed sequences via the input mask MNM_N, which samples the signal at intervals of θ\theta and holds it constant over one virtual-node interval, before being applied to modulate the laser’s driving current. The prediction performance is assessed in terms of the Normalized Root Mean Square Error (NRMSE) calculated between the outputs of the linear readout and the ground-truth target sequence.

Lorenz System

The Lorenz system is a canonical three-dimensional chaotic flow governed by

x˙=σ(y−x),y˙=x(ρ−z)−y,z˙=xy−βz(6)\dot{x}=\sigma(y-x),\quad\dot{y}=x(\rho-z)-y,\quad\dot{z}=xy-\beta z \tag*{(6)}

with the standard parameters σ=10\sigma=10, ρ=28\rho=28, and β=8/3\beta=8/3. The system is integrated with a fourth-order Runge–Kutta scheme using a step size of Δtint=0.001\Delta t_{\mathrm{int}}=0.001, and the trajectory is sampled every 0.02 time units. The xx-component is linearly normalized to [−1,1][-1,1]. The reservoir is tasked with predicting the state two steps ahead, corresponding to a physical horizon of 0.04. Compared with the MG task, the Lorenz system involves three coupled degrees of freedom and a more complex attractor topology, imposing a stronger demand on the reservoir’s nonlinear mapping capacity.

Santa Fe Laser Intensity Dataset

We use Dataset A from the 1994 Santa Fe Time Series Prediction Competition [24, 25], which records the intensity of a far-infrared laser operating in a chaotic regime. The distributed record contains 10093 consecutive samples; the sequence is linearly normalized to [−1,1][-1,1] without further preprocessing. The reservoir is tasked with predicting the intensity two sampling steps ahead, where one step corresponds to one consecutive data point. Because the experimental record does not specify a physical sampling interval, the prediction horizon is expressed in units of data points rather than in physical time. Because the dataset originates from an experimental single-mode laser system, it provides a realistic test of the reservoir’s ability to handle noise and non-ideal dynamics.

Numerical Simulation and System Parameters

The TDRC system is modelled using the complete Maxwell-Bloch equations with a 4th-order Runge-Kutta (RK4) integration scheme. Spontaneous emission noise is added as an additive white Gaussian term after each RK4 step [26]. To isolate performance differences caused by internal laser physics, the external macroscopic delay τ\tau was kept identical across all classes (50 units of TphT_{\mathrm{ph}}). A fixed integration step dt=0.002dt = 0.002 was uniformly adopted across all laser classes, which remains sufficiently small to guarantee numerical stability and precise temporal alignment even for the fastest internal dynamics (e.g., T2=0.01T_{2} = 0.01). The intrinsic physical parameters governing the laser dynamics and the numerical integration settings are summarized in Table I.

ParameterSymbolClass AClass BClass CA_AClass CB_B
Polarization lifetimeT2T_20.010.012.02.0
Carrier lifetimeTT0.01392.02.0392.0
Normalized cavity decay ratecc(2δωT2+2)2+1\left(\frac{2\delta\omega}{T_2+2}\right)^2+1(2δωT2+2)2+1\left(\frac{2\delta\omega}{T_2+2}\right)^2+1(2δωT2+2)2+1\left(\frac{2\delta\omega}{T_2+2}\right)^2+1(2δωT2+2)2+1\left(\frac{2\delta\omega}{T_2+2}\right)^2+1
Physical feedback delayτ\tau50.050.050.050.0
Number of virtual nodesNvN_v20202020
Integration time stepdtdt0.0020.0020.0020.002
Integration scheme4th-order Runge-Kutta4th-order Runge-Kutta4th-order Runge-Kutta4th-order Runge-Kutta

Table I. Intrinsic physical parameters and numerical simulation settings for the four laser classes. All time values are normalized to the cavity photon lifetime TphT_{\mathrm{ph}}. The parameter cc ensures the first laser threshold at Ppump=0.5P_{\mathrm{pump}} = 0.5. The optimized control parameters (κ,Pbase,θ,γ,δω,ϕ)(\kappa, P_{\mathrm{base}}, \theta, \gamma, \delta\omega, \phi) are summarized separately in Table 2. The intermediate quantum-dot-like Class AQ_Q (T2=0.01T_2 = 0.01, T=2.0T = 2.0) is documented in the Supplementary Information.

Fair Comparison Methodology

To guarantee a fair comparison among the four laser classes, we adhere to three methodological principles. First, steady-state intensity normalization: the pump bias PbaseP_{\mathrm{base}} for each class is uniformly set to Pbase=1.95P_{\mathrm{base}} = 1.95, establishing a consistent baseline signal-to-noise ratio. Second, individual parameter optimization: the node spacing θ\theta and input scaling factor γ\gamma are separately scanned and optimized for each class via high-resolution NRMSE grid searches, so that the subsequent comparison captures the best achievable performance of each laser class rather than the performance evaluated at an arbitrarily chosen shared parameter set. Third, unified data protocol: for all three benchmarks, the first 3,000 samples are used for training and the subsequent 5,000 for testing, with a washout of 100 samples at the beginning of each sequence.

To avoid temporal degeneracy and to reflect realistic reservoir operation, the number of virtual nodes is set to Nv=20N_{\mathrm{v}} = 20 for all classes. Then, to understand the dynamical origin of the computational differences among the four laser classes, we investigate the intrinsic temporal responses of each laser type to a transient pump perturbation. The distinct response characteristics, from rapid monotonic recovery in Class A, damped relaxation oscillations in Class B, to high-frequency coherent ringing enabled by the finite polarization lifetime in Class CAC_{\mathrm{A}} and Class CBC_{\mathrm{B}} lasers, are shown in the Supplementary Material (Supplementary Fig. S1). These underlying dynamical responses fundamentally shape the fading memory and nonlinearity available for the TDRC tasks.

Results and Discussion

Physical Phase Diagrams and Topological Time-Scale Signatures

To investigate how the interplay between polarization and carrier dynamics governs the operational phase space, we carry out extensive parameter sweeps and use Permutation Entropy (PE) to quantify dynamical complexity [27]. As shown in Figure 2, the PE mappings for the four laser classes are mapped across two distinct parameter spaces: the feedback strength–detuning landscape (δω,κ)(\delta\omega, \kappa) (Figure 2a–Figure 2d) and the feedback phase–detuning landscape (δω,ϕ)(\delta\omega, \phi) (Figure 2e–Figure 2h). In these color maps, deep purple regions indicate low entropy (stable continuous-wave operation or highly regular periodic limit cycles), while the transition toward green and yellow represents elevated entropy, indicative of complex chaotic dynamics.

Permutation Entropy mappings of the four laser classes

Figure 2. Permutation Entropy (PE) mappings of the four laser classes in two distinct parameter spaces: (a–d) the feedback strength–detuning plane (δω,κ)(\delta\omega, \kappa) at fixed phase ϕ=π/2\phi= \pi/2 and (e–h) the feedback phase–detuning plane (δω,ϕ)(\delta\omega, \phi) at fixed feedback strength κ=0.05\kappa= 0.05. Panels (a, e) represent Class A; (b, f) Class B; (c, g) Class CA_A; and (d, h) Class CB_B. The white cross in every panel marks the unified physical baseline (δω=1.5,κ=0.05,ϕ=π/2)(\delta\omega= 1.5, \kappa= 0.05, \phi= \pi/2) used for all task evaluations in this work; its position is (1.5,0.05)(1.5, 0.05) in the (δω,κ)(\delta\omega, \kappa) plane and (1.5,π/2)(1.5, \pi/2) in the (δω,ϕ)(\delta\omega, \phi) plane.

We first examine the (δω,κ)(\delta\omega, \kappa) parameter space. Class A and Class CAC_{\mathrm{A}} exhibit smooth, parabolic continuous-wave (CW) stability basins, characteristic of strongly damped systems where fast relaxation suppresses long-term memory accumulation [28, 29]. Class B (Figure 2b), by contrast, reveals a highly structured network of multistable locking domains, a well-known signature of the intricate bifurcation structure in delayed feedback lasers [17, 30, 31]. This deterministic structural confinement, arising from the beating between relaxation oscillations and the external cavity frequencies, which are spaced by 1/τ1/\tau [30], restricts the edge-of-stability dynamics to narrow parameter islands [4]. In contrast, Class CBC_{\mathrm{B}} (Figure 2d) demonstrates the topological smoothing power of a finite polarization lifetime. Acting as an intrinsic low-pass filter [32, 33], it effectively dissolves the sharp locking boundaries of Class B into a broadened, continuous, and parametrically resilient operational basin.

We next evaluate the (δω,ϕ)(\delta\omega, \phi) parameter space. As shown in Figure 2e and Figure 2g, both Class A and Class CAC_{\mathrm{A}} display vertically uniform, low-entropy backgrounds, which in both cases stems from the suppression of phase interference due to either fast damping (Class A) or short-term memory (Class CAC_{\mathrm{A}}).

Class B (Figure 2f), by contrast, displays a striking pattern of diagonally striated, discrete “short-bar-like” structures, which are governed by the phase-matching condition δω⋅τ+ϕ=2mπ\delta\omega\cdot\tau+ \phi= 2m\pi [34]. In contrast, Class CBC_{\mathrm{B}} (Figure 2h) smears these discrete bars into continuous, broadened diagonal bands, once again demonstrating the smoothing effect of its finite polarization lifetime as a low-pass filter [35].

These distinct topologies, as illustrated in Supplementary fig:S1, are fundamentally rooted in the lasers’ internal transient dynamics and validate our central thesis: the slow polarization of Class CB_B acts as a topological smoother, dissolving the sharp bifurcation boundaries and phase sensitivity observed in the strongly damped and long-memory regimes, thus unifying hardware-robust stability with the rich dynamics required for photonic neuromorphic processors.

Based on these diagrams, we choose a unified detuning of δω=1.5\delta\omega= 1.5 for all subsequent task evaluations. This choice is motivated by three considerations. First, it lies in a region where all four laser classes exhibit non-trivial dynamics—neither deep in the stable continuous-wave regime nor beyond the edge of stability—thereby allowing each class to operate near its optimal computational regime. Second, it is sufficiently far from the dense multistable locking domains of Class B to avoid the fragile ECM locking that would otherwise dominate, while still preserving the topological smoothing effect in Class CB_B. Third, it provides a common ground for fair comparison, ensuring that the observed performance differences arise from the intrinsic time-scale hierarchies rather than from arbitrary detuning choices. This detuning is therefore adopted as the unified physical baseline in Table II and used throughout the task evaluations. This baseline is marked by a white cross in every panel of Figure 2. The subsequent optimization of task-specific parameters (θ,γ)(\theta, \gamma) is then conducted on this common physical foundation.

Laser ClassUnified Physical BaselineOptimized Engineering Params.
κ\kappaPbaseP_{\mathrm{base}}δω\delta\omegaϕ\phiθopt\theta_{\mathrm{opt}}γopt\gamma_{\mathrm{opt}}
Class A0.051.951.5π/2\pi/21.301.39
Class B0.051.951.5π/2\pi/237.110.28
Class CA_{\mathrm{A}}0.051.951.5π/2\pi/21.861.22
Class CB_{\mathrm{B}}0.051.951.5π/2\pi/218.571.12

Table II. Summary of the unified physical baseline and the optimized control parameters derived from the NRMSE scans.

Optimal Parameter Mapping Guided by Task-Driven Performance

To establish an unbiased physical platform for comparing the four laser classes, we first map the global predictive performance across the pump-feedback parameter space (PbaseP_{\mathrm{base}}, κ\kappa), with a detailed phenomenological analysis provided in the Supplementary Materials (Fig. S2). Briefly, the error landscapes reveal that the ultra-long carrier memory of Class B restricts stable operation to narrow parameter corridors bounded by horizontal resonant bands, while Class CA_{\mathrm{A}} suffers from localized optical quenching (i.e., gain truncation) under simultaneous high-pump and strong-feedback driving. In contrast, Class CB_{\mathrm{B}} utilizes its finite polarization lifetime (T2=2.0T_2 = 2.0) as a dynamic low-pass filter to maintain an expansive, remarkably uniform stability plateau. Guided by these macroscopic topological boundaries, we establish a universally stable baseline at Pbase=1.95P_{\mathrm{base}} = 1.95 and κ=0.05\kappa= 0.05 (weak-feedback regime). This unified operating point safely navigates the dynamical constraints of Class B and the over-saturation limits of Class CA_{\mathrm{A}}, ensuring that the subsequent optimization of task-specific information encoding parameters (virtual node interval θ\theta and input scaling factor γ\gamma) is conducted on a rigorously equitable physical foundation. It is necessary to mention that this baseline defines only the intrinsic laser physics; the task-specific information-encoding parameters (θ\theta and γ\gamma) are independently optimized for each class in a two-dimensional grid search to capture their respective maximum computational capacity. This systematic optimization strategy follows established best practices for photonic RC parameter tuning and TDRC operating-point selection [13, 14, 36].

We perform high-resolution two-dimensional (θ\theta, γ\gamma) scans for each laser class under the unified physical baseline. The resulting optimal operating points are summarized in Table II. The 2D grid search locates physically implementable optima for all classes, with γ\gamma remaining within a normal driving range (0.28≤γ≤1.390.28 \le\gamma\le1.39). This range corresponds to the standard weak-to-moderate modulation regime in which the laser operates below the gain-truncation boundary, ensuring that the optimized operation points are compatible with physically realizable experimental conditions.

The resulting optimization landscapes exhibit distinct physical signatures directly reflecting each laser’s temporal-scale hierarchy. For Class A (Figure 3a, Figure 3e), the γ\gamma-scan at their respective optimal θ\theta reveals smooth, broad valleys with the minimum located at γ≈1.4\gamma\approx1.4. This operating point works for all the 4 lasers and they do not transition into chaotic behaviour. The θ\theta-scan exhibits a broad but shallow valley, confirming the absence of an intrinsic resonance and consistent with the memoryless nature of these classes.

High-resolution 1D parameter sweeps for the four primary laser classes

Figure 3. High-resolution 1D parameter sweeps for the four primary laser classes. Top row (a–d): NRMSE as a function of the input scaling factor γ\gamma at each class’s optimal node interval θopt\theta_{\mathrm{opt}}. Bottom row (e–h): NRMSE as a function of the virtual node interval θ\theta at each class’s optimal γopt\gamma_{\mathrm{opt}}. Columns correspond to Class A (a, e), Class B (b, f), Class CA_{\mathrm{A}} (c, g), and Class CB_{\mathrm{B}} (d, h). Each panel shows a single class, color-coded as follows: Class A in red, Class B in green, Class CA_{\mathrm{A}} in blue, and Class CB_{\mathrm{B}} in purple. The red P-shaped markers denote the globally optimal parameters yielding the minimum NRMSE. In panels (e) and (h), the dashed vertical line marks the boundary between the main scan and the high-resolution supplementary scan for Class A and Class CB_{\mathrm{B}}.

For the Class B lasers (Figure 3b, Figure 3f), the θ\theta-scan reveals a pronounced V-shaped valley, with an optimum located at θopt=37.11\theta_{\mathrm{opt}} = 37.11. This significantly longer node interval, corresponding to θopt/τro≈1.32\theta_{\mathrm{opt}}/\tau_{\mathrm{ro}} \approx1.32, allows the ultra-long carrier lifetime (T=392.0T = 392.0) to accumulate temporal correlations without degenerating into a static-mapping regime. The corresponding γ\gamma-scan is dominated by an extremely shallow optimum at γopt=0.28\gamma_{\mathrm{opt}} = 0.28, indicating that Class B requires a very weak input coupling to avoid over-driving its slow carrier reservoir.

For Class CA_{\mathrm{A}} (Figure 3c, Figure 3g), with T2=T=2.0T_2 = T = 2.0, the optimal operating point is at θopt=1.86\theta_{\mathrm{opt}} = 1.86 and γopt=1.22\gamma_{\mathrm{opt}} = 1.22. The corresponding ratio θopt/τro≈0.93\theta_{\mathrm{opt}}/\tau_{\mathrm{ro}} \approx0.93 places this laser squarely within the classical dynamic TDRC regime. Both the γ\gamma-scan and the θ\theta-scan exhibit smooth, well-defined parabolic minima, confirming that Class CA_{\mathrm{A}} operates at a genuinely balanced operating point where both carrier and polarization dynamics contribute constructively.

In contrast to Class B, the Class CB_{\mathrm{B}} laser (Figure 3d, Figure 3h) forms a distinctly different landscape. Its θ\theta-scan (obtained by combining the main scan θ∈[20,50]\theta\in[20, 50] with a high-resolution supplementary scan θ∈[10,20]\theta\in[10, 20]) exhibits a minimum value around θopt=18.57\theta_{\mathrm{opt}} = 18.57 (θopt/τro≈0.66\theta_{\mathrm{opt}}/\tau_{\mathrm{ro}} \approx0.66). This pronounced sensitivity to θ\theta detuning is a direct manifestation of the slow polarization lifetime (T2=2.0T_2 = 2.0). By acting as an inline temporal filter, polarization dynamics smooth out high-frequency inter-node timing mismatches, which shifts the optimal node spacing relative to Class B. The γ\gamma-scan locates the optimum at γopt=1.12\gamma_{\mathrm{opt}} = 1.12, which is comparable in magnitude to that of Class CA_{\mathrm{A}}, demonstrating that the polarization-induced low-pass filtering operates at a fundamentally similar driving regime as the perfectly matched Class CA_{\mathrm{A}}.

These optimization results demonstrate that engineering parameters cannot be chosen arbitrarily. They must be precisely aligned with the intrinsic relaxation frequencies and dynamic damping rates of the underlying physical substrate. In particular, the physically implementable operating points identified here (all with γ≤1.4\gamma\le1.4) remain well below the gain-truncation boundary that limits the memoryless classes. The perfect time-scale matching of Class CA_{\mathrm{A}} provides the most energy-efficient operation, and the low-pass filtering of Class CB_{\mathrm{B}} offers the broadest operational tolerance. Table II summarizes the unified physical baseline alongside the optimal control parameters across all four laser classes.

High-Precision Prediction and Multi-Mask Robustness

Having established the optimal physical baseline and engineering parameters (Table II), we evaluate the high-precision predictive performance and stochastic encoding robustness across the four primary laser classes. The reservoir models were trained on 3,000 consecutive data points and evaluated on an independent out-of-sample test set comprising 5,000 time steps of the Mackey-Glass chaotic time series. To quantitatively benchmark input encoding resilience, we performed 20 independent evaluation trials for each laser class, employing a newly generated, distinct random continuous uniform in each iteration.

Figure 4 summarizes the multi-mask prediction performance. Figure 4a shows a representative 800-step prediction trajectory for Class A. The thin black curve is the true target, the red dashed curve is the ensemble average over 20 independent random masks, and the shaded region denotes the ±1σ\pm1\sigma. The prediction closely follows the target, and the narrow ±1σ\pm1\sigma band indicates small mask-to-mask variance, confirming that the optimized parameter configuration provides the complex transient response required for this task.

Multi-mask robustness plots

Figure 4. Multi-mask robustness. (a) Representative 800-step prediction trajectory. The thin black curve is the true target, the thick red dashed curve is the ensemble average over 20 independent random masks, and the shaded region indicates the ±1σ\pm1\sigma band. The prediction closely follows the target, and the narrow ±1σ\pm1\sigma band indicates small mask-to-mask variance. (b) Statistical distributions of the NRMSE over the 20 mask realizations for all four laser classes. Boxes show the interquartile range, black lines the median, white diamonds the mean, and colored dots the individual mask realizations. The similar box widths indicate that the mask-induced variance is comparable across classes, while the different box positions reflect the distinct intrinsic nonlinearity of each laser class. The performance differences reported in Table III therefore originate from the intrinsic laser physics rather than from mask randomness.

Figure 4b compares the NRMSE distributions across the 20 mask realizations for all four laser classes. The distributions are tight and of comparable width for all classes (standard deviations between 0.001 and 0.002), indicating that the mask-to-mask variation is consistent across classes and does not alter their relative ranking. The mean NRMSE values, however, differ systematically: Class A achieves 0.010, Class CA_{\mathrm{A}} 0.011, Class CB_{\mathrm{B}} 0.014, and Class B 0.021. These differences reflect the distinct intrinsic nonlinearity and time-scale hierarchy of each class: the instantaneous response of Class A supports a stable static nonlinear mapping; the perfectly matched time scales of Class CA_{\mathrm{A}} yield the richest intermediateorder nonlinear response; the finite polarization lifetime of Class CB_B acts as an intrinsic low-pass filter that slightly smooths high-frequency mask components; and the ultra-long carrier lifetime of Class B limits its response speed. The performance differences reported in Table III therefore originate from the intrinsic laser physics rather than from mask randomness.

Laser ClassMackey-GlassLorenzSanta Fe (Dataset A)
Class A0.0103 ±\pm 0.00140.0567 ±\pm 0.01480.5233 ±\pm 0.1631
Class B0.0217 ±\pm 0.00310.1544 ±\pm 0.04000.6955 ±\pm 0.0426
Class CA_A0.0105 ±\pm 0.00240.0506 ±\pm 0.01420.4251 ±\pm 0.0651
Class CB_B0.0141 ±\pm 0.00130.0763 ±\pm 0.03210.5341 ±\pm 0.0650

Table III. Task Generalization Performance (NRMSE ±\pm Standard Deviation) across different benchmark tasks. All values are the mean ±\pm standard deviation over 10 independent random masks.

The ability to sustain low prediction error with minimal variance across stochastic mask realizations provides an essential physical foundation for practical, hardware-ready neuromorphic processors [37].

Task Generalization and Information Processing Capacity Decomposition

To assess whether the observed performance trends are specific to a single benchmark or reflect a more general task dependence, we evaluate the four primary laser classes against two additional, distinct chaotic benchmarks: the three-dimensional Lorenz system and the experimental Santa Fe laser intensity dataset. All evaluations are performed under the unified physical baseline and class-optimized control parameters established in Table II. For each task, the readout weights are retrained on the corresponding training set. Table III summarizes the cross-task prediction performance, reported as the mean Normalized Root Mean Square Error (NRMSE) alongside its standard deviation (±1σ\pm1\sigma) across 10 independent stochastic mask realizations.

The empirical benchmarks reveal a strongly task-dependent performance landscape, shaped by the interplay between driving strength and intrinsic time scales. On the short-horizon Mackey–Glass task, Class A (0.0103) and Class CA_A (0.0105) achieve comparable, and the lowest, NRMSE—their difference is well within one standard deviation—while Class A exhibits a smaller mask-to-mask spread (0.0014 vs. 0.0024). This result is consistent with the operation of both classes in the static-nonlinear-mapping regime: their high content of instantaneous response is ideally suited to this essentially short-memory task. However, as the task demands higher-order nonlinear complexity over a longer effective horizon, the matched time scales of Class CA_A (θopt/τσ≈0.93\theta_{\mathrm{opt}}/\tau_{\sigma} \approx0.93) become decisive, and Class CA_A overtakes Class A on both the Lorenz (0.0506 vs. 0.0567) and the Santa Fe (0.4251 vs. 0.5233) benchmarks. The advantage of Class CA_A is particularly pronounced on the experimental Santa Fe dataset, where it leads Class A by roughly 20%, whereas the same comparison on the deterministic Lorenz system yields a narrower margin of about 11%. This distinction suggests that real-world data, with its non-stationarity and measurement noise, rewards the balanced time-scale hierarchy of Class CA_A even more strongly than idealized chaotic flows. Notably, the mask-to-mask spread of Class A on Santa Fe (±0.1631\pm0.1631) is more than twice that of Class CA_A (±0.0651\pm0.0651), indicating that the matched time scales of Class CA_A also provide a substantially more robust operating point for noisy experimental data.

Class B exhibits the poorest performance across all three benchmarks (MG: 0.0217, Lorenz: 0.1544, Santa Fe: 0.6955). The ultra-long carrier lifetime (T=392.0T = 392.0) combined with the polarization-free dynamics makes the system respond too slowly for the rapidly varying Lorenz and Santa Fe attractors, and the weak input coupling required to stabilize the slow carrier reservoir limits the amount of information that can be encoded.

In contrast, Class CB_B occupies an intermediate position, outperforming Class B by 35% on Mackey–Glass (0.0141 vs 0.0217), 51% on Lorenz (0.0763 vs 0.1544), and 23% on Santa Fe (0.5341 vs 0.6955). The polarization-mediated low-pass filtering plays a decisive role: by smoothing the high-frequency components of the deterministic input mask and enabling a stronger input coupling (ϑopt=1.12\vartheta_{\mathrm{opt}} = 1.12 vs. 0.28 for Class B), Class CB_B simultaneously achieves a larger effective modulation depth and a richer nonlinear feature space. The task-dependent improvement is largest for the Lorenz benchmark, which combines a longer prediction horizon with a deterministic, noise-free attractor; on the experimental Santa Fe dataset, the additional noise partially offsets the benefit of the polarization filter, and the improvement over Class B is correspondingly smaller. On Santa Fe, Class CB_B (0.5341) and Class A (0.5233) realizations. are statistically indistinguishable, as their difference is much smaller than the standard deviation of either class; nevertheless, Class CB_B retains a systematic advantage over Class B on all three benchmarks, confirming the role of the finite polarization lifetime as a nonlinear-capacity booster.

These results highlight a fundamental trade-off. Short-horizon nonlinear mapping favors fast carriers (Class A), while long-horizon chaotic prediction requires the balanced time scales of Class CA_A. The optimal architecture is therefore not a fixed hardware choice but a task-adaptive design decision.

It is important to note that the absolute NRMSE values reported in Table III are not directly comparable to those in the literature. Most photonic TDRC demonstrations of Mackey–Glass prediction employ substantially more virtual nodes (typically Nν≥50N_{\nu} \ge50, often several hundred) and different prediction horizons [4, 13, 14, 38]. In contrast, the present work uses only Nν=20N_{\nu} = 20 and a relatively short horizon of 0.1 for Mackey–Glass. Under these more constrained conditions, the NRMSE values of 0.010–0.022 for Mackey–Glass and 0.051–0.154 for Lorenz fall within the range reported for photonic reservoir computing platforms. The absolute NRMSE on the Santa Fe dataset (0.425–0.696) is substantially larger, as expected for a real experimental record with non-stationary dynamics and measurement noise; values in the range 0.3–0.6 are typical for this benchmark in the literature. Across all three tasks, the central result of this work is not the absolute accuracy but the relative ranking of the four laser classes, which remains consistent with the IPC decomposition presented below.

To systematically unravel the physical origin of these variations, we employ the Information Processing Capacity (IPC) framework based on orthogonal Legendre polynomial expansions. The IPC metric, originally developed by Dambre et al. [39] and further refined with finite-sample bias correction by Kubota et al., provides a task-independent measure of a system’s linear memory and nonlinear computational capability [40]. The degree-decomposed IPC spectrum provides a quantitative mapping of linear memory versus higher-order nonlinear transformation capacity. Recent studies have successfully demonstrated the effectiveness of this decomposition in designing optimized photonic RC systems [6, 23].

The IPC decomposition (Figure 5) reveals a nuanced hierarchy that, when interpreted together with the benchmark results, quantitatively explains the task-dependent reversal observed in Table III. The four laser classes display markedly different linear vs. nonlinear capacity balances.

Stacked bar chart showing IPC decomposition across four laser classes

Figure 5. Information Processing Capacity (IPC) decomposition across the four primary laser classes, evaluated at each class’s optimal control parameters (see Table II). The stacked bars show the spectral distribution of computational capacity across Legendre polynomial degrees from d=1d = 1 (linear memory) to d=6d = 6 (high-order nonlinearity). The numbers above each bar (ΣC\Sigma C) indicate the total capacity, i.e., the sum over all degrees d=1d = 1–66. The x-axis labels the four laser classes; the y-axis is the IPC contribution.

The Class A laser achieves the highest linear IPC (4.70), reflecting its instantaneous response. Its nonlinear IPC (12.87) is substantial but lower than that of Class CA_A. This specific balance – high linear, moderate nonlinear – perfectly matches the requirements of short-horizon Mackey–Glass prediction.

The Class B laser exhibits the lowest total capacity across both linear (1.99) and nonlinear (4.11) components. The ultra-long carrier lifetime severely restricts the achievable number of independent polynomial mappings, consistent with the systematically poor performance of this class on all three benchmarks.

Class CA_A achieves the highest nonlinear IPC (16.28) and the second-highest total capacity (19.29). The dominant nonlinear contribution reflects the capacity of the well-matched time scales to generate a rich hierarchy of higher-order Legendre features — precisely the resource that the Lorenz and Santa Fe tasks require. Class CBC_B displays a substantial nonlinear IPC (9.78) that is 2.4× larger than that of Class B, while its linear IPC (2.03) remains comparable. This selective enhancement of nonlinear capacity directly underpins the 52% improvement over Class B on the Lorenz and Santa Fe tasks. Notably, the polarization filter acts purely as a nonlinear-capacity booster; it does not affect the linear component, which is why Class CBC_B’s advantage over Class B is much larger for the higher-dimensional chaotic benchmarks than for the short-horizon Mackey-Glass task.

The correspondence between IPC and benchmark performance is striking. The ranking of the four classes by linear IPC (A > CAC_A > CBC_B > B) closely matches the ranking by Mackey-Glass accuracy, with Class A and Class CAC_A statistically indistinguishable. The ranking by nonlinear IPC (CAC_A > A > CBC_B > B) matches the Lorenz ranking exactly and the Santa Fe ranking to within the mask-to-mask uncertainty, where Class A and Class CBC_B are statistically indistinguishable. This agreement demonstrates that the IPC framework provides a rigorous physical basis for understanding the task-dependent performance of the laser-based TDRC systems.

These generalization benchmarks, directly illuminated by IPC spectral decomposition, confirm that carrier and polarization relaxation lifetimes do not merely dictate total capacity magnitudes; they fundamentally sculpt the system’s intrinsic computational spectrum. The optimal architecture must achieve a balance between memory depth and responsiveness that matches the specific dynamical demands of the target task. In the present benchmark suite, this balance is best realized by Class CAC_A for the nonlinear chaotic tasks and by Class A for the short-horizon prediction task, whereas the extreme underdamping of Class B and the (partial) over-smoothing of Class CBC_B lead to systematically reduced performance on one or more benchmarks. By tailoring physical relaxation timescales – and specifically by engineering the correct separation between T2T_2 and TT – photonic neuromorphic processors can be custom-engineered to match the topological demands of specific complex time-series tasks, establishing timescale engineering as a core design paradigm for hardware-adaptive optical computing.

Impact of Spontaneous Emission Noise on Reservoir Stability

To evaluate physical scalability under realistic quantum noise, we stress the optimized TDRC systems with varying spontaneous emission factors βnoise\beta_{\mathrm{noise}} (10−610^{-6} to 10−110^{-1}). Figure 6 presents both the prediction accuracy (a) and the intrinsic coherence degradation (b), revealing a critical threshold at βnoise≈10−3\beta_{\mathrm{noise}} \approx10^{-3}. Below the noise threshold, all four laser classes exhibit deterministic behavior. Beyond this value, the intrinsic carrier lifetime TT dominates the diverging fates.

Robustness of the optimized TDRC systems against intrinsic spontaneous emission noise

Figure 6. Robustness of the optimized TDRC systems against intrinsic spontaneous emission noise. (a) Average NRMSE for the Mackey-Glass one-step prediction task as a function of the spontaneous emission factor βnoise\beta_{\mathrm{noise}}, spanning 10−610^{-6} to 10−110^{-1} on a logarithmic scale. Each curve corresponds to one laser class: Class A (red), Class B (green), Class CAC_A (blue), and Class CBC_B (purple). Error bars indicate the standard deviation across 10 independent mask realizations. The dashed gray horizontal line marks the engineering target threshold of NRMSE = 0.1. (b) Corresponding second-order intensity correlation g(2)(0)g^{(2)}(0) of the free-running lasers (without input modulation). The dashed gray line marks the Poissonian (coherent) limit g(2)(0)=1g^{(2)}(0) = 1; the dotted gray line marks the thermal limit g(2)(0)=2g^{(2)}(0) = 2. Class A exhibits the slowest degradation of coherence, reaching only g(2)(0)=1.43g^{(2)}(0) = 1.43 at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}, whereas the long-carrier classes (Class B and Class CBC_B) approach the thermal limit more rapidly. The finite polarization lifetime T2=2.0T_2 = 2.0 in Class CBC_B provides a modest but measurable protection compared with Class B, as reflected in both panels. Note that the NRMSE values in panel (a) increase smoothly and monotonically for all four classes, without any sharp threshold-like transition.

Class A (red) displays the highest resilience to quantum noise, with NRMSE rising only to 0.47 even at the extreme limit βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}. The exceptionally fast carrier dynamics (T=0.01T = 0.01) act as a broadband noise sink: spontaneous emission fluctuations are dissipated on a timescale much shorter than the symbol period, so they are naturally filtered out before the next virtual node is sampled. This is directly confirmed by the corresponding intensity correlation, which remains close to the coherent limit (g(2)(0)=1.43g^{(2)}(0) = 1.43 at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}, the lowest among all classes).

Class CA_A (blue) exhibits the second-best noise resilience (NRMSE = 0.58 at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}), with a smooth monotonic increase as a function of βnoise\beta_{\mathrm{noise}} and no evidence of a threshold-induced ”coherence avalanche.” The finite polarization lifetime (T2=2.0T_2 = 2.0) provides a mild additional filtering effect, but the dominant noise-robustness mechanism remains the finite carrier lifetime (T=2.0T = 2.0), which dissipates fluctuations more quickly than the long-carrier classes.

Class CB_B (purple) achieves intermediate noise resilience (NRMSE = 0.78 at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}), but a residual T2T_2-mediated benefit is clearly visible. Despite sharing the same ultra-long carrier lifetime as Class B (T=392T = 392), Class CB_B outperforms Class B by 20% at the extreme noise limit (0.78 vs. 0.98). The finite polarization lifetime (T2=2.0T_2 = 2.0) acts as an inline low-pass filter that partially blocks high-frequency spontaneous emission before it can be integrated by the long-lived carriers. This effect is manifested in the intensity correlation, which reaches 1.87 for Class CB_B versus 1.94 for Class B at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}.

Class B (green) displays the poorest noise resilience (NRMSE = 0.98 at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}), acting as an unbuffered noise integrator. The ultra-long carrier lifetime (T=392.0T = 392.0) causes spontaneous emission fluctuations to be integrated over very long timescales, driving the intensity correlation towards the thermal limit (g(2)(0)→2.0g^{(2)}(0) \to2.0). The absence of any intrinsic filtering mechanism means the noise accumulates monotonically with the pump modulation.

These stress tests establish a fundamental physical trade-off for photonic neuromorphic design. The dominant determinant of noise resilience is the carrier lifetime: short-carrier classes naturally dissipate noise faster than they can be integrated, whereas long-carrier classes suffer from progressive noise accumulation. Against this dominant background, a finite polarization lifetime T2T_2 provides a systematic but secondary (approximately 20%) enhancement by acting as an inline low-pass filter, which is most valuable precisely when the carrier lifetime is long enough for the polarization filter to be the only available protection mechanism. The design principle that emerges from these results is therefore not simply “retain a finite T2T_2,” but rather “engineer the polarization lifetime in concert with the carrier lifetime to achieve the specific time-scale hierarchy demanded by the target application.”

Conclusion and Discussion

In conclusion, we have established time-scale engineering as a nuanced design paradigm for photonic reservoir computing. Our unified theoretical framework reveals that retaining a finite polarization lifetime T2T_2 fundamentally reshapes the phase-space topology, information processing capacity, and noise immunity of TDRC systems.

This comprehensive study exposes a fundamental task-dependent reversal that defines the design space of photonic neuromorphic hardware:

(i) Short-horizon nonlinear mapping favors fast carriers. Class A and Class CA_A achieve comparable, and the lowest, NRMSE on the short-horizon Mackey-Glass task (0.0103 and 0.0105, respectively); the difference is within one standard deviation, consistent with their operation in the static-nonlinear-mapping regime. Class A retains the highest linear IPC (4.70) and the smaller mask-to-mask spread.

(ii) Long-horizon chaotic prediction favors matched time scales. Class CA_A achieves the best Lorenz and Santa Fe performance (0.0506 and 0.4251), consistent with its high nonlinear IPC (16.28) and its balanced time-scale hierarchy (τopt/τ0≈0.93\tau_{\mathrm{opt}}/\tau_0 \approx0.93, corresponding to a clock-cycle-to-delay ratio of Nϑopt/τ≈0.74N\vartheta_{\mathrm{opt}}/\tau\approx0.74).

(iii) The long-carrier classes exhibit systematically reduced performance across all three tasks. Class B is the poorest performer (MG: 0.0217, Lorenz: 0.1544, Santa Fe: 0.6955), and Class CB_B retains a residual advantage over Class B (MG: 0.0141, Lorenz: 0.0763, Santa Fe: 0.5341) due to polarization-induced low-pass filtering, but neither can compete with Class A or Class CA_A. On the Santa Fe dataset, Class A (0.5233) and Class CB_B (0.5341) are statistically indistinguishable. We note that the fixed delay τ=50\tau= 50 used throughout this study corresponds to τ/T≈0.128\tau/T \approx0.128 for the long-carrier classes, i.e., only about one-eighth of their carrier lifetime; this fixed short delay specifically limits the memory available to slow-carrier dynamics. A proportionally longer delay would be required for a fully fair comparison of these classes, which we leave for future work.

(iv) Under realistic quantum noise, the carrier lifetime TT is the dominant determinant of noise resilience. Class A retains the lowest NRMSE (0.47) even at βnoise=10−1\beta_{\mathrm{noise}} = 10^{-1}, whereas Class B degrades to 0.98. A residual T2T_2-mediated benefit of approximately 20% persists in the slow-carrier Class CB_B (0.78 vs. 0.98).

By mapping abstract relaxation dynamics to actionable device-level parameters (e.g., spin-relaxation rates in VCSELs or photon lifetimes in high-β\beta microcavities), our findings provide concrete physical design principles for engineering fault-tolerant, high-throughput optical computing chips. We emphasize that the deliberate retention of a finite polarization lifetime is a powerful, practical lever, but that its benefit is maximized only when the polarization lifetime is engineered in the correct relationship to the carrier lifetime and to the task-specific time-scale demands.

Acknowledgements

TW thanks the support from National Natural Science Foundation of China (Grant No. 62475206), and Key Research and Development Plan of Shaanxi Province of China (Grant No. 2024GH-ZDXM-42). KL acknowledges support from the European Union’s Horizon 2020 programme under grant agreement number 101129904, SPIKEPro.

References

  1. [1]S. R. Ahmed, R. Baghdadi, M. Bernadskiy, N. Bowman, R. Braid, J. Carr, C. Chen, P. Ciccarella, M. Cole, J. Cooke, et al., Nature 640, 368 (2025).
  2. [2]B. J. Shastri, A. N. Tait, T. Ferreira de Lima, W. H. Pernice, H. Bhaskaran, C. D. Wright, and P. R. Prucnal, Nature Photonics 15, 102 (2021).
  3. [3]M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, and J. Sun, Nature Communications 15, 2056 (2024).
  4. [4]L. Appeltant, M. C. Soriano, G. Van der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso, and I. Fischer, Nature Communications 2, 468 (2011).
  5. [5]G. Tanaka, T. Yamane, J. B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Neural Networks 115, 100 (2019).
  6. [6]J. Huang, T. Wang, K. Lüdge, Y. Han, S. Xiang, and Y. Hao, Neural Networks 197, 108383 (2026), ISSN 0893-6080.
  7. [7]H. Hasegawa, K. Kanno, and A. Uchida, Nanophotonics 12, 869 (2023).
  8. [8]A. Röhm, L. Jaurigue, and K. Lüdge, IEEE Journal of Selected Topics in Quantum Electronics 26, 1 (2020).
  9. [9]I. Estébañez, A. Argyris, and I. Fischer, Optics Letters 48, 2449 (2023).
  10. [10]T. Wang, C. Jiang, Q. Fang, X. Guo, Y. Zhang, C. Jin, and S. Xiang, Progress in Electromagnetics Research 178, 1 (2023).
  11. [11]K. Kanno, A. A. Haya, and A. Uchida, Optics Express 30, 34218 (2022).
  12. [12]T. Wang, H. Zhou, Q. Fang, Y. Han, X. Guo, Y. Zhang, C. Qian, H. Chen, S. Barland, S. Xiang, et al., Chaos, Solitons & Fractals 181, 114673 (2024).
  13. [13]D. Brunner, M. C. Soriano, C. R. Mirasso, and I. Fischer, Nature Communications 4, 1364 (2013).
  14. [14]H. Dong, L. Jaurigue, and K. Lüdge, Physica Status Solidi (RRL) – Rapid Research Letters 19, 2400433 (2025).
  15. [15]J. R. Tredicce, F. T. Arecchi, G. L. Lippi, and G. P. Puccioni, Journal of the Optical Society of America B 2, 173 (1985).
  16. [16]A. Roos, S. Meinecke, and K. Lüdge, Scientific Reports 11, 18558 (2021).
  17. [17]S. Wieczorek, B. Krauskopf, T. B. Simpson, and D. Lenstra, Physics Reports 416, 1 (2005).
  18. [18]K. Lüdge and B. Lingnau, in Synergetics (Springer, 2020), pp. 31–47.
  19. [19]R. Aust, T. Kaul, C. Ning, B. Lingnau, and K. Lüdge, Optical and Quantum Electronics 48, 109 (2016).
  20. [20]B. Lingnau, J. Turnwald, and K. Lüdge, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 377, 20180124 (2019).
  21. [21]F. T. Arecchi, G. L. Lippi, G. P. Puccioni, and J. R. Tredicce, Optics Communications 51, 308 (1984), ISSN 0030-4018.
  22. [22]H. Deng, G. L. Lippi, J. Mørk, J. Wiersig, and S. Reitzenstein, Advanced Optical Materials 9, 2100415 (2021).
  23. [23]T. Hülser, F. Köster, K. Lüdge, and L. Jaurigue, Nanophotonics 12, 937 (2023).
  24. [24]U. Hübner, N. B. Abraham, and C. O. Weiss, Physical Review A 40, 6354 (1989).
  25. [25]A. S. Weigend and N. A. Gershenfeld, eds., Time Series Prediction: Forecasting the Future and Understanding the Past, vol. 15 of Santa Fe Institute Studies in the Sciences of Complexity (Addison-Wesley, Reading, MA, 1994), ISBN 0-201-62601-2.
  26. [26]K. Pyragas and T. Pyragienė, Physical Review E 112, 015301 (2025).
  27. [27]J. Cancio, C. Masoller, and M. Marconi, Physical Review Letters 135, 093802 (2025).
  28. [28]M. Virte, K. Panajotov, H. Thienpont, and M. Sciamanna, Nature Photonics 7, 60 (2013).
  29. [29]N. Takemura, M. Takiguchi, and M. Notomi, Journal of the Optical Society of America B 38, 699 (2021).
  30. [30]M. Marconi, C. Métayer, A. Acciavatti, M. Boyer, A. Gomel, T. Quiniou, C. Masoller, M. Giudici, and J. R. Tredicce, Physical Review Letters 125, 134102 (2020).
  31. [31]M. Sciamanna and K. A. Shore, Nature Photonics 9, 151 (2015).
  32. [32]H. Haken, in Light and Matter Ic/Licht und Materie Ic (Springer, 1984), pp. 1–304.
  33. [33]G. H. Van Tartwijk and G. P. Agrawal, Progress in Quantum Electronics 22, 43 (1998).
  34. [34]H.-J. Wünsche, S. Bauer, J. Kreissl, O. Ushakov, N. Korneyev, F. Henneberger, E. Wille, H. Erzgräber, M. Peil, W. Elsäßer, et al., Physical Review Letters 94, 163901 (2005).
  35. [35]B. Romeira, J. M. Figueiredo, and J. Javaloyes, Chaos: An Interdisciplinary Journal of Nonlinear Science 27 (2017).
  36. [36]H. Jaeger and H. Haas, Science 304, 78 (2004).
  37. [37]M. Goldmann, F. Köster, K. Lüdge, and S. Yanchuk, Chaos: An Interdisciplinary Journal of Nonlinear Science 30 (2020).
  38. [38]L. Larger, A. Baylón-Fuentes, R. Martinenghi, V. S. Udaltשov, Y. K. Chembo, and M. Jacquot, Physical Review X 7, 011015 (2017).
  39. [39]J. Dambre, D. Verstraeten, B. Schrauwen, and S. Massar, Scientific Reports 2, 514 (2012).
  40. [40]T. Kubota, H. Takahashi, and K. Nakajima, Physical Review 3, 043135 (2021).

Paper details

vlnc:W_4a4duwjz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents