The increasing demand for efficient, high-speed, and low-power information processing systems has driven substantial interest in neuromorphic computing [1,2]. Inspired by the brain’s ability to process information through sparse spiking activity, spiking neural networks (SNNs) provide a biologically plausible and energy-efficient paradigm for tackling complex computational tasks [3]. Notable neuromorphic hardware examples implemented with traditional digital electronic technologies include SpiNNaker by the University of Manchester [4,5], IBM’s TrueNorth [6], Intel’s Loihi [7], and BrainScaleS by the University of Heidelberg [8]. Despite their promise, the electronic hardware realization of SNNs presents challenges, particularly in achieving scalability, robustness, and ultra-fast signal processing on digital and analog platforms. Addressing these limitations while handling the complexity of bio-inspired spiking mechanisms remains a key research focus [9].

Neuromorphic photonics offers a potential solution by combining neural-inspired computation with the high speed and bandwidth of light signals [10–12]. However, many implementations still rely on analog (non-spiking) optical signals, which can suffer from noise, limited robustness, and reduced interpretability. Among photonic platforms, silicon microring resonators (MRRs) have emerged as promising candidates due to their compactness, low-power operation, ease to integrate in reduced-footprint photonic chips, and compatibility with CMOS technologies [13]. MRRs feature linear filtering capabilities for tunable synaptic weights [14–16] and nonlinear passive processes like two-photon absorption (TPA) and free-carrier dispersions (FCDs) [17] enabling useful nonlinear transformation for photonic neural networks [18–24]. The competition between free-carrier and thermal effects in high-quality-factor (≈104) MRRs, also enables passive key neuron functionalities such as continuous self-pulsation [14,15] and deterministic spiking dynamics [25,26]. Thanks to the self-pulsation dynamics, the optical memory of a single MRR can extend beyond the typical timescales of free-carrier (1–45 ns) and thermal (60–280 ns) effects. Indeed, optical perturbations (relative to the self-pulsation periodicity) can reproducibly delay the emission of the next spike, enabling long memory effects [27] that prove effective, for example, in tackling prediction tasks [22]. Moreover, these perturbations can alter the shape of subsequent spiking patterns in complex MRR structures, depending on the input pulse rate, thus enabling input rate classification [27]. Deterministic spiking dynamics, in contrast to self-pulsation, have been achieved using a pump-and-probe approach to controllably elicit an optical spike.

This work focuses on the achievement and understanding of deterministic spiking dynamics in MRRs. As a first contribution, we demonstrate the generation of optical spikes in MRRs using a single modulated carrier signal. This eliminates the need for a separate continuous wave (CW) pump [25,26], simplifying the architecture and improving energy efficiency. In this configuration, unlike the pump-and-probe method, the refractory period that follows the spike emission depends on the pulse duration, with the longest periods observed when the optical spike is allowed to fully develop according to the nonlinear dynamics of the MRR. We also demonstrate the ability of MRRs to integrate multiple wavelength-multiplexed optical input signals, firing all-optical spiking patterns in response. A direct neuromorphic application demonstrated experimentally in this work is high-speed coincidence detection. To achieve this functionality, used widely by biological neurons (for example, in the auditory cortex to locate sound procedence), we leverage the wavelength division multiplexing (WDM) optical integration capability of MRRs to detect two synchronous sub-threshold optical pulses at different wavelengths generating an output optical spike within a fast coincidence detection window (ns-rate). This prompt and high-contrast spiking response is enabled by the use of negative, rather than positive, input wavelength detuning with respect to the cold MRR resonance.

These properties are then combined to develop a framework for sparse photonic spiking reservoir computing [28,29]. The system leverages the optical MRR’s spiking behavior and a novel simultaneous WDM and time division multiplexing (TDM) encoding mechanism proposed in this work, to process information faster and with robust optical spiking outputs. Its performance is validated on the Iris-Flower dataset classification task, where the system achieves classification test accuracies up to 92% using a very simple information encoding mechanism and operation with only 48 optical spiking reservoir nodes, demonstrating remarkably sparse activity, with an average of just three spikes per flower datapoint’s classification. Additionally, during generalization to test datasets, the photonic MRR spiking reservoir exhibits an accuracy drop of only 2% compared to the 12% drop observed when using the corresponding analog input, employing the same number of trained parameters and training samples. Given the simplicity of the spiking system and its predisposition for binary output weights, we believe the proposed framework for passive, sparse, spiking all-optical reservoir computing holds significant promise when extended in the future to more complex structures of coupled spiking MRRs [30], providing guidance for the design of future passive MRR-based photonic spiking neural networks.

The manuscript is organized as follows. In Section 2 we first characterize the optical spiking regimes that can be achieved in the MRR under study when subject to both CW and pulsed optical input signals. In Section 3 we then demonstrate the operation of a silicon MRR as an optical spiking neuron able to perform first coincidence detection between two WDM-encoded optical input signals, and second as an all-optical WDM-TDM photonic spiking reservoir computer. Finally, in Sections 4 and 5 we draw the future outlook and conclusions of the work, respectively.

The experimental setup designed to excite optical spiking responses in a silicon MRR under both single- and multi-wavelength input operation is illustrated in Fig. 1. Two optical input lines carry the information to be processed by the MRR, each encoded onto different wavelength carriers. In each line, light is first generated by a tunable laser source (TLS1,2) and amplitude is modulated by a Mach–Zehnder modulator (MZM). The latter is driven by an RF signal generated by an arbitrary waveform generator (AWG) (MOKU:Pro) and synchronized with the MZM in the other optical line using a trigger signal from another AWG instrument. Light is then amplified by an erbium-doped fiber amplifier (EDFA). Each optical line also includes two polarization controllers (PCs) to maximize polarization coupling in order to optimize the operation of the MZMs and the coupling of the injected light into the silicon chip (TE polarization). The average optical power in each line is adjusted by varying the transmission (T1, T2) of variable optical attenuators (VOAs). Note that tuning the two input VOA transmissions in the range [0,1] enables to compare the MRR’s optical response under single-wavelength (T1=0 or T2=0) and WDM ([T1,T2]≠[0,0]) injection. The two optical lines are then combined in a 2×2 50:50 fiber directional coupler, with one of its outputs attenuated (Ti) and then directed to a fast photodiode (PDi) to monitor the encoded input optical waveform, while the other coupler’s output feeds the MRR’s optical input port. Each optical channel provides a maximum average optical power of 4.31 mW, when T1=1 and T2=0, or T2=1 and T1=0 at the microring waveguide, after accounting for the 3 dB losses from the grating coupler. Ti is calibrated to not saturate PDi when [T1,T2]=[1,1]. An optical fiber array is used to couple the optical input/output light with the photonic chip, via the input, through and drop ports of the integrated MRR. The through signal is measured with an optical spectrum analyzer (OSA) and also serves for light-alignment purposes. The drop signal carrying the optical spiking responses of the silicon MRR is coupled out of the chip with the fiber array, detected by an amplified photodetector (PD, Thorlabs PDA8GS), recorded with a 16 GHz bandwidth oscilloscope (OSC, Rohde& Scwartz RTP084), and stored in a computer for analysis. To maintain the stability of the detuning between the pumps and the cold MRR resonance wavelengths, the chip temperature is actively regulated using a proportional-integral-derivative (PID) controller connected to a Peltier cell and a 10kΩ thermistor. This limits resonance wavelength fluctuations to within 3 pm over a 5 h monitoring period, comparable to the 2 pm wavelength stability of the input lasers.

The silicon MRR under study is in an add-drop configuration with a waveguide cross-section of 220nm×500nm and radius r=6.75μm, and is point-like coupled to the two bus waveguides with a gap of 238 nm. Its linear spectrum shown in Fig. 2(a) is acquired by feeding the EDFA’s amplified spontaneous emission (ASE) into the MRR’s input port and analyzing the through transmission of the MRR using the OSA in Fig. 1. Input wavelengths belonging to the ASE spectrum that are resonant with the MRR are routed to its drop port, hence emerging as dips in the spectrum of Fig. 2(a), separated by a 14.412 nm free spectral range. Two working resonances in Fig. 2(a) are selected and used in this work. We couple light from tunable laser 1 (TLS1) with a pump wavelength λ1,p into the MRR’s resonance centered at λ1,c=1545.37nm (full width at half-maximum, FWHM = 83 pm) with a cold detuning defined as Δλ1=λ1,p−λ1,c. Similarly, light from tunable laser 2 (TLS2), with a pump wavelength λ2,p, is optically coupled into the MRR’s resonance centered at λ2,c=1558.782nm (FWHM = 67 pm), with a cold detuning referred to as Δλ2=λ2,p−λ2,c.

In MRRs designed with high quality factors (Q≈104), resonant light can trigger nonlinear two-photon absorption effects within the MRR’s waveguide. This process generates additional free carriers in the conduction band and phonons within the waveguide. The interplay between free-carrier effects and temperature variation shapes the optical power output of the MRR, resulting in a characteristic spike firing dynamical profile [31]. To analyze the nonlinear optical spiking responses of the MRR under injection of a CW optical signal, we examined various combinations of wavelength detuning and optical input power. Each selected resonance in Fig. 2(a) is interrogated using the corresponding optical line from Fig. 1, under different configurations of cold detuning (Δλ1 or Δλ2) and input optical power (adjusted via the VOA transparencies: T1 or T2). The resulting optical self-pulsing maps (i.e., the self-pulsing rate dependence on the input parameters) are presented in Fig. 2(b), illustrating the response for individual MRR resonance scans. Due to thermally induced resonance shifts, self-pulsing predominantly occurs at positive wavelength detunings, extending beyond the cold resonance width of the MRR. While low negative detunings can also support self-pulsing, its rate is lower [as shown in the common color bar in Fig. 2(c)] and requires higher input optical power. Therefore, in these passive MRRs the threshold input optical power required for TPA-induced spiking depends significantly on the input laser wavelength. When both optical lines are simultaneously activated, the self-pulsing region expands further into positive and negative detunings, highlighting the MRR’s optical wavelength integration capability [see Fig. 2(c)]. In this scenario, both optical lines are operated with identical detunings (Δλ1=Δλ2=Δλ1,2) and input optical powers (T1=T2=T1,2). In this way, clear optical self-pulsing responses are exhibited at the drop port of the MRR, as it is shown in Fig. 2(d) for three different cases of cold detuning: Δλ1,2=[−90,−70,80]pm with T1,2=0.75. At Δλ1,2=−90pm the MRR remains quiescent, whereas continuous optical spiking dynamics with increasing firing rates emerge for Δλ1,2=−70pm and Δλ1,2=80pm.

Controlled TPA-based optical spikes have been demonstrated in silicon MRRs using a pump-and-probe technique [25,26]. In this method, a continuous wave (CW) pump signal maintains the MRR on the onset of the spiking threshold, while a radio-frequency (RF) probe signal encodes low-amplitude pulses to selectively trigger spiking responses. Notably, each spike emission is followed by a refractory period during which the MRR cannot respond to new stimuli. In this work, we investigate a novel approach to elicit deterministic spiking events in a silicon MRR, in which the device is subject to a single optical (pulsating) injection. In Fig. 3 we analyze the MRR’s response to optical pulses of varying durations and amplitudes for a selected negative wavelength detuning [Δλ1,Δλ2]=[−70,−70]pm. As in Fig. 2(b), the values of Δλ1 and Δλ2 are configured so they represent identical wavelength detunings for both optical lines, with respect to the two consecutive cold resonances of the MRR targeted for injection. Orange and green lines correspond to conditions where only optical line 1 (T2=0) or optical line 2 (T1=0) is open, respectively. Purple lines depict the MRR output optical spiking responses when both channels are simultaneously open and aligned. The first column in Fig. 3 shows the MRR’s response to the injection of a long optical pulse, emulating the beginning of a CW optical pump condition. The pulse amplitude is sufficient to generate spikes when either channel 1 (orange line) or channel 2 (green line) is open individually. Interestingly, more spikes are produced in channel 2 despite their identical optical input pulse amplitudes, as the coupled optical power into the MRR at that wavelength of λ2,c=1558.782nm is higher than the lower wavelength of the orange resonance at λ1,c=1545.37nm [see Fig. 2(b)]. When both optical channels are open (purple line), the number of optical spikes fired by the MRR further increases, indicating the system’s ability to integrate the intensities from the optical input signals injected at both wavelength channels. These initial plots provide a reference for estimating the duration of an optical spike, which is used to calibrate the optical pulse sequence in the subsequent tests (central panels in Fig. 3). Here, each optical pulse has a duration of 200 ns, allowing the spike to fully evolve, thus ensuring the maximum competition time between free-carrier and thermal effects before thermal shifts initiate the refractory period. The presence of a refractory period is confirmed by the MRR’s inhibition to respond to subsequent optical input pulses (200 ns apart) until it fully recovers and can elicit a new optical spike. In contrast, when the optical pulse duration is shortened to 50 ns (right panels in Fig. 3), the spike formation is prematurely interrupted. As the pulse ends, free carriers decay rapidly, halting phonon generation and resulting in a smaller red-shift of the resonance compared to complete spike generation. Consequently, the refractory period is significantly reduced or bypassed entirely. This is evident as the MRR produces now a spike with each injected optical pulse, despite their higher amplitude. These results demonstrate that optical spikes with refractory periods can be induced in silicon MRRs without a pump-and-probe method, relying solely on the energy of modulated wavelength carriers. The refractory period becomes dependent on the pulse duration, with the longest periods observed when the optical spike forms entirely according to the MRR nonlinear dynamics. This dependency, whose quantitative analysis is reported in Appendix A, highlights a trade-off inherent to the single-carrier approach integrating over a single optical pulse: unlike the pump-and-probe method, where the refractory period is decoupled from the probe pulse duration, using a single carrier links the spike dynamics directly to the optical pulse characteristics. This difference underscores the distinctive behavior and constraints introduced by relying solely on a single modulated wavelength carrier.

In Appendix B we illustrate also the MRR’s response to the same input patterns under positive detunings [Δλ1,Δλ2]=[+80,+80]pm. In this case the optical spikes are delayed by preliminary thermal effects, show a lower spike-to-quiescence ratio, and any refractory period, if present, is likely much shorter than the pulse intervals used in these measurements.

In biological neuronal networks, action potentials are fired when the converging inputs to a neuron raise its membrane potential above a characteristic threshold level. Sub-threshold inputs can trigger a spike if they arrive synchronously and their combined energies exceed the activation threshold of the neuron. If their arrival times are not sufficiently close, the membrane potential remains below the activation threshold. This process, known as coincidence detection, plays a crucial role in reducing jitter caused by spontaneous activity and in other functions such as sound localization [32,33]. Optical spiking neurons based upon silicon MRRs can leverage wavelength division multiplexing (WDM) to achieve coincidence detection, allowing optical inputs at different wavelengths—even within the same physical channel—to be processed simultaneously. This enhances the ability to detect synchronous information across both physical and wavelength channels. To test the WDM-based optical coincidence detection capability we injected two optical pulses at different wavelengths with controlled delays between them. These pulses were detuned relative to the two consecutive cold resonance positions shown in Fig. 2(a) by [Δλ1,Δλ2]=[−70,−70]pm. Each sub-threshold pulse had a duration of 200 ns and, when applied individually or with significant delay, did not elicit an optical spike (first and last columns of Fig. 4). However, when the two optical pulses (at different wavelengths) overlapped in time, their combined energy was sufficient to trigger an optical spike in the MRR, demonstrating coincidence detection operation within a fast (ns-rate) temporal window. Notably, coincidence detection relies on a prompt optical spiking response from the MRR. For this reason, we used negative wavelength detunings, which enable faster spike firing responses in the system. Positive detunings, in contrast, would introduce delays due to the thermal transient required for the MRR to lock with the input signals (see Fig. 9 in Appendix B). This delay would necessitate longer input pulses, ultimately slowing down the optical processing.

The deterministic excitation of optical spikes in an MRR using single-wavelength perturbations—without requiring a pump-and-probe scheme—and the integration of multi-wavelength sub-threshold perturbations (via WDM coincidence detection) are here combined to propose a new framework for an on-chip, passive photonic spiking reservoir computer (RC). This builds on advances in photonic time-multiplexed RC, where the reservoir state unfolds in the dynamical evolution of a physical nonlinear node, which is sampled at different time instants to form a set of NV time-multiplexed virtual nodes. This computing paradigm was first demonstrated with analog non-spiking nonlinear nodes [34] and recently also demonstrated with a spiking vertical-cavity surface emitting laser (VCSEL) neuron [35]. Here, we go beyond the state-of-the-art by extending this concept to an optical spiking silicon MRR neuron excited via two-photon absorption physics, hence delivering a hardware-friendly yet extremely versatile photonic spectro-temporal SNN that allows fully passive operation and on-chip integration.

The network is evaluated on the well-known Iris-Flower dataset classification task, a standard benchmark in machine learning [36]. This dataset contains 150 flowers from three Iris-Flower species—Setosa, Versicolar, and Virginica—each described by four analog features (sepal and petal lengths and widths). We leverage the two optical wavelength-multiplexed carriers applied in the setup in Fig. 1 to propose a novel encoding mechanism that simultaneously encodes two features per channel and therefore doubles the processing speed. As depicted in Fig. 5(a), each normalized feature, relative to the corresponding feature across all flowers in the dataset, is multiplied by a mask signal M(t) containing Nm=24 analog mask values, producing a waveform with Nv=48 analog values per channel. The duration of each mask value is carefully set to 200 ns, which aligns closely with the duration of an MRR optical spike (see Fig. 3). This ensures a direct one-to-one correspondence between an output virtual node and a triggered spike. Given this tangible physical basis, we will henceforth refer to virtual nodes as “spiking nodes.” The two input waveforms are synchronized and wavelength-multiplexed into a single optical line using a fiber directional coupler (Fig. 1). One-half of the multiplexed optical input signal is split for power monitoring at PDi [optical input waveforms in Fig. 5(a)], while the other half is directed to the MRR’s input port. The corresponding drop-port optical waveform [purple drop in Fig. 5(a)] is recorded by an oscilloscope, from which the optical spiking vector is extracted. The spiking vector assigns a “0” for non-spiking nodes and a “1” for nodes where the drop signal exceeds the spike threshold [indicated by the horizontal black line in Fig. 5(a)]. The digital spiking vector represents the state of the photonic spiking reservoir for each flower sample, as depicted in the network diagram in Fig. 5(b). To introduce temporal coupling between consecutive reservoir nodes and enhance the diversity of spiking patterns representing different flowers in the dataset, the input optical signals are detuned by Δλ1,2=−70pm. This wavelength detuning ensures that each 200 ns mask modulation generates optical spikes followed by a refractory period (see Fig. 3), influencing future node states. The digital spiking nodes in the reservoir are linearly combined into a three-dimensional analog output vector using three sets of output weights Wout. A winner-takes-all scheme labels the node with the largest value in the output vector as the predicted class of the input flower. During the training phase, the output weights are optimized using a ridge regression algorithm, which minimizes the error between the predicted and target classes. Once trained, the network with the optimized Wout is evaluated on a test dataset to assess the generalization of the training process to previously unseen samples. Training and testing accuracies are calculated as the proportion of correctly classified flower datapoints relative to the total number in the respective datasets. The scoring procedure follows a Monte Carlo cross-validation approach, as follows. The MRR’s optical response to all 150 flowers is acquired 10 times. After each acquisition, the dataset is randomly split into training (120 samples) and test (30 samples) subsets, 100 times, generating 100 training and testing scores. This procedure, repeated for all 10 acquisitions, produces two final vectors—one for training and one for testing—each containing 1000 values. The mean values and standard deviations of these vectors represent the final training and testing accuracies, providing a robust estimate of the network’s performance across multiple randomized trials. Notably, these scores reflect real-time processing, as no average is applied a priori either by the oscilloscope or through post-processing to smooth the acquired optical traces before the training.

The system achieves an optimal test accuracy of 92%±4% on the Iris-Flower dataset classification task using just 48 spiking reservoir nodes. With 200 ns/node, this yields a processing time per datapoint of 9.6 μs. Achieving this performance requires a correct selection of the optical spiking threshold, a critical parameter for converting the photonic spiking patterns from the MRR into digital spike vectors. As illustrated in Fig. 6(a), the spike threshold should not be too low to allow the MRR response to surpass the threshold even in the absence of spike firing events, leading to the detection of false spikes (top panel), nor too high to exceed the maximum MRR optical spike excursions, resulting in no spikes being detected (bottom panel). Ideally, the threshold should be positioned between these extremes (central panel) to ensure the detection of only optical spikes elicited in the MRR. As a natural consequence, the system’s performance on the Iris-Flower task is highly sensitive to the chosen spike threshold. This dependency is shown in Fig. 6(b). Here the test accuracy (purple line, left y-axis) is plotted across a range of optical spike thresholds (0–140 mV), the latter referring to voltage values coming from the read-out after photo-detection. The figure also shows the average number of detected optical spikes per flower (blue line, right y-axis), with error bars representing the minimum and maximum number of spikes observed across the dataset. At a spike threshold of zero, 48 digital spikes are detected, indicating that the MRR response exceeds the threshold in all 48 nodes for every flower response. This lack of differentiation between flower responses leads to the minimum test accuracy. When the spike threshold is increased, fewer reservoir nodes are converted into digital spike representations, leading to an initial improvement in test accuracy. However, for spiking thresholds above 45 mV, the number of spikes quickly drops to zero, causing a sharp decline in test accuracy. The highest accuracy of 92% is achieved within the threshold range of 25–45 mV, where only true optical spikes fired by the MRR are detected. This optimal range reflects the high optical spike-to-quiescent ratio enabled by negative wavelength detuning optical injection, demonstrating the robustness of silicon MRRs in photonic spike processing. Interestingly, in this range, the average number of optical spikes per flower reduces to just three (out of 48 reservoir nodes), which coincides with the number of classes in the task. This sparse spiking activity (three out of 48 nodes) is combined to the variability in which nodes are activated within the 48-node spike vector to give a diverse set of spike patterns to represent different Iris-Flower datapoints and achieve the system’s high accuracy. For example, Fig. 6(c) illustrates the spike patterns obtained for all flowers at an optimal threshold of 40 mV. Flower datapoints from the first class primarily activate optical spikes in the first half of the nodes (0–24), while flowers in the second and third classes exhibit sparse spiking activity across all 48 nodes, with some nodes activating uniquely in specific cases. This sparse activity underlines the way the MRR’s refractory period (enabled by masked pulses lasting 200 ns and negatively detuned) couples nodes in the spiking vector. The confusion matrix in Fig. 6(d) indicates that the system classifies Iris Setosa, Versicolar, and Virginica flowers with an average test accuracy of 97%, 91%, and 88%, respectively. These results were obtained with both input optical wavelength-multiplexed channels simultaneously activated with a negative detuning of Δλ1,2=−70pm, and an optical power calibrated at T1,2=0.5 transparency for the VOA in the setup (≈2.15mW average optical power within the MRR input waveguide, per channel).

The system was also evaluated on the Iris-Flower dataset classification task across a range of input optical powers, adjusted via the VOA transparencies T1 and T2. Figure 7(a) presents the training and test accuracies under three different configurations: with a single optical channel open and injected into the MRR (λ1, yellow; or λ2, green) and with both channels open and injected simultaneously into the system (λ1+λ2, purple). The highest accuracy (0.92±0.04) is achieved when both wavelength-multiplexed channels are open at [T1,T2]=[0.5,0.5], as reported previously. Using only one optical wavelength channel results in reduced performance due to incomplete information transfer, as each channel encodes only two of the four features of each Iris-Flower datapoint. Specifically, for VOA transparencies of 0.5, the performance is the lowest when only the λ1 channel is open, improves when only the λ2 is open, and peaks when both channels are active simultaneously. This behavior is explained in Fig. 7(b), left panels, which display the MRR’s optical input (blue lines) and drop response for a single Iris-Flower datapoint at a VOA transparency of 0.5. With only λ1 open, no optical spikes are generated, resulting in a zero spiking vector for all flower datapoints. With only λ2 open, a single spike is generated, enabling some differentiation between flowers and improving accuracy. When both channels are open, the produced optical spike pattern becomes richer, emphasizing the TPA-based integration capability of silicon MRRs, and leading to the optimal performance. Accordingly, Fig. 7(a) shows that accuracies begin to improve once the power of the single channel or both channels combined is sufficient to generate optical spikes. At higher input power levels, such as T1=1 and T2=1 [see Fig. 7(b), right panels], the number of optical spikes triggered in the MRR increases. Interestingly, however, the performance degrades compared to the optimal configuration at [T1,T2]=[0.5,0.5]. This degradation occurs because higher power causes more nodes to spike uniformly, reducing pattern diversity and consequently lowering accuracy. The time traces corresponding to the optimal configuration in Fig. 7(b), left panels, suggest that the most diverse and effective spike patterns are obtained when the MRR is near the onset of the spike firing activation, before both channels are opened simultaneously.

For comparison, Fig. 7(a) also presents the accuracy achieved by applying ridge regression directly to the input signal recorded by PDi when both optical lines, λ1 and λ2, are open (blue line). In this case, it is important to highlight that the state of each virtual node is analog (not spiking) and corresponds to the average detected power within the respective mask time-slot. The training accuracy [Fig. 7(a), left panel] indicates that 48 analog (real-valued) virtual nodes are sufficient to overfit the optical input data, resulting in a notable generalization drop in test accuracy [Fig. 7(a), right panel]. In contrast, the nonlinear transformation of the signal into a (photonic) spiking pattern helps mitigate overfitting. For instance, at [T1,T2]=[0.5,0.5], the spiking MRR reservoir achieves higher accuracy than the analog input, with only a ≈2% drop from training to testing datasets (purple line). In comparison, the analog input signal exhibits a larger ≈12% accuracy drop. Additionally, Fig. 7(a), right panel, reveals that the standard deviation of accuracy over 1000 scores decreases with increasing input optical power (T1, T2) for the analog input signal, whereas it remains nearly constant for the spiking MRR reservoir. This distinction highlights the inherent robustness of spiking nodes against optical input noise, particularly at lower optical powers, making them promising for energy-efficient computation.

Silicon microring resonators have already shown significant potential for analog photonic neural networks [18–24]. For example, a single microring operated at the edge of its stable and self-pulsing regime, leveraging free-carrier and thermal nonlinearities, achieved an accuracy of (99.3±0.2)% on the Iris-Flower dataset classification task using 50 analog nodes [19]. Similarly, a photonic reservoir composed of silicon microring arrays operating in the linear regime and scattering light to a nonlinear camera demonstrated (97.2±0.2)% accuracy with just 18 analog pixel nodes [16].

The results presented in this work demonstrate that a single silicon MRR is also highly effective when excited in its optical spiking dynamics, achieving an accuracy of (92±4)% with 48 reservoir nodes, of which only three are, on average, spiking. Notably, these results were obtained without any averaging of the optical traces (either through the oscilloscope or post-processing across multiple acquisitions) and without the use of EDFA or optical filters at the chip’s output. As a comparison, in Ref. [35] a recently reported spiking reservoir computer using a VCSEL neuron achieved 91.7% accuracy using 512 reservoir nodes (while only 48 were used in the present work) and >97% average accuracy when configured with a larger network of 1024 nodes. Indeed, ideally, just three spiking nodes activated selectively for the three flower classes would suffice to solve the task, and larger reservoir sizes via longer input mask patterns [M(t)] would enhance the probability to diversify the spiking patterns and find these three nodes. This highlights the simplicity and robustness of this photonic spiking reservoir approach, demonstrating its scalability potentials to permit the handling of more complex classification tasks involving a greater number of classes. Notably, it also paves the way for implementing digital output weights that activate in conjunction with the output spikes of interest. This represents a significant simplification compared to analog output weights, simplifying the on-chip implementation of the reservoir’s output layer and online learning algorithms used for its training.

Increasing the number of mask nodes M(t) enhances the diversity of optical spiking patterns but comes at the cost of reduced processing speed. A promising alternative is to design advanced architectures based on coupled MRRs to generate more diverse spiking patterns while minimizing the number of mask nodes and maintaining high processing speeds [27,30]. Additionally, the processing speed could be significantly improved by fully leveraging the novel WDM-based encoding method introduced in this work, using more than two wavelength-multiplexed optical carriers. This would couple all the individual datapoint features in separate wavelength channels, instead of pairing them up (see Fig. 5), enriching the MRR optical spiking dynamics and potentially performance. The processing speed can also be optimized by tuning the nonlinear timescales that govern the MRR’s spiking dynamics. Typical free-carrier and thermal timescales in silicon MRRs range around 1–45 ns and 60–280 ns, respectively [37,38]; but these could be further reduced by embedding, for instance, the waveguide in the MRR within a p-i-n junction [39].

We experimentally demonstrated fast, deterministic optical spiking responses in integrated silicon microring resonators. Notably, these are achieved under single optical injection detuned negatively with respect to the microring cold resonance wavelengths, without the need for pump-and-probe methods. Within this scenario the input pulse duration becomes a critical parameter, with long refractory periods following the spike emission only when the latter forms entirely before the input pulse is turned off. This capability highlights the microring’s ability to integrate sub-threshold inputs across multiple wavelength-multiplexed optical channels, enabling more efficient, high-speed processing without necessarily requiring a pump signal.

These findings fuel the demonstration of a novel information encoding scheme allowing both wavelength and time division multiplexing to encode and launch information into a spiking microring reservoir computer. The encoding scheme is properly set with optical carriers at negative wavelength detunings and with timescales that allow the microring to enter the refractory period. In this way, we show experimentally that the silicon microring can deliver fast, sparse, and high-contrast optical spike patterns, achieving a high average classification accuracy (92%) on the Iris-Flower dataset classification task. Notably, this result relies on real-time traces with only three sparse optical spikes per flower datapoint (in average), out of just 48 spiking reservoir nodes and without the need of any averaging procedure during read-out or post-processing. Looking forward, the scalability of this system can be extended by integrating additional wavelength-multiplexed channels, expanding the temporal length of the mask signal, and exploring architectures of on-chip coupled silicon microrings. All these enhance the richness of the optical spiking responses. Notably, this framework enables compatibility with digital output weights, offering a great simplification for their on-chip integration and learning protocols directly in optical hardware. The capability of silicon microrings to meaningfully convert analog information inputs into sparse optical spike trains, added to their recently demonstrated self-pulsing long-term memory effects [27], has broad implications for applications in light-enabled edge-computing, sensing, and other real-time processing scenarios, shining promising for future fast, compact, and efficient photonic spiking neural networks.

Acknowledgment. We gratefully thank Dr. Alessio Lugnan for useful suggestions and interesting discussions. The authors also acknowledge funding support from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme, which has now ended, and during which the chip used in our experiments was fabricated.