Fig. 1. Sketch of the experimental setup probing the MRR’s spiking response under WDM operation. TLS: tunable laser source; PC: polarization controller; MZM: Mach–Zehnder modulator; EDFA: erbium doped fiber amplifier; VOA: variable optical attenuator having transmission T; PS: power source; AWG: arbitrary waveform generator; MOKU: AWG model; PD: photodetector; OSA: optical spectrum analyzer; OSC: oscilloscope; COMP: computer. Optical lines are indicated in yellow, green (single wavelength), and purple (WDM), while blue lines indicate electrical connections.

Fig. 2. Experimental characterization of the MRR under study. (a) Optical spectra measured at the through port of the MRR under the injection of amplified spontaneous emission of an EDFA. (b) Closer look at the two resonances selected in this work, with corresponding self-pulsing maps, with the spike rate highlighted in color bar for each combination of wavelength detuning (Δλ) and CW optical power (T, VOA transmission). The maximum optical power of each channel (T1=1, T2=1) within the microring input waveguide, after 3 dB grating losses, is 4.31 mW. (c) Self-pulsing map achieved when both wavelength channels are open with the same detuning (Δλ1,2) and power (T1,2). (d) Examples of optical traces at the drop port for specific Δλ1,2-T1,2 in (c). The color bar common to all self-pulsing maps is provided in (c).

Fig. 3. Refractory period enabled in MRRs without a pump-and-probe approach. Spiking response patterns obtained for input sequences of optical pulse (blue lines) with diverse durations: several μs (left), 200 ns (center), and 50 ns (right), interleaved by 200 ns in all cases. The detuning operation is set to [Δλ1,Δλ2]=[−70,−70]pm. Yellow, green, and purple lines indicate the configurations where λ1 only (yellow), λ2 only (green), or both λ1 and λ2 channels (purple) are open. Input pulse sufficiently long to accommodate the complete spiking generation [competition between free-carrier (FC) + thermal (TH) effects], enabling the MRR’s refractory period (center). For shorter pulses the refractory period is bypassed (right).

Fig. 4. WDM coincidence detection enabled in silicon MRRs. Two optical sub-threshold pulses coupled at different MRR resonance wavelengths (with detuning Δλ1,2=−70pm) are injected with diverse delays into the MRR, eliciting a spike response when they overlap.

Fig. 5. (a) WDM proposed encoding scheme. The four analog flower’s features are split and time-multiplexed into two wavelength channels (orange and green lines), according to a mask signal M(t) and then input to the MRR reservoir. The corresponding spiking response from the MRR drop port (purple line) is encoded into a binary spike vector of zeros and ones (blue circles), according to a spike threshold (black horizontal line), and linearly combined to classify the flower. (b) Corresponding spiking reservoir computing neural network based on an MRR and wavelength-time multiplexing.

Fig. 6. Iris-Flower task results obtained by the spiking MRR reservoir when using negative detunings Δλ1,2=−70pm and input optical transparencies T1,2=0.55. (a) Variation of the detected spiking nodes across three levels of the spike threshold. (b) Test accuracy dependence on the spike threshold, with corresponding average number of spikes per flower, while using 48 digital nodes; the best test accuracy is 0.92±0.04. (c) Spike patterns generated by all flowers and (d) confusion matrix, both achieved for a spike threshold of 40 mV and averaging across 10 MRRs’ response acquisitions.

Fig. 7. Performance in the Iris-Flower task obtained by the spiking MRR reservoir when using negative detunings Δλ1,2=−70nm and 48 reservoir spiking nodes. (a) Training (left) and test (right) accuracies dependence on the input optical transparencies of the two wavelength channels T1 and T2. Different color lines indicate configurations where only a single input wavelength channel is open (λ1, yellow; λ2, green) or both channels are open simultaneously (λ1+λ2, purple and blue). Error bars indicate the standard deviation over 1000 scores. (b) The same flower encoded with only one wavelength channel open (orange and green curves) or two channels open simultaneously (purple curves) at VOA transparencies of 0.5 (left) and 1.0 (right), with the corresponding MRR’s spiking responses. Blue lines indicate the relative optical input signals.

Fig. 8. MRR refractory period as a function of the optical pulse duration, under single laser injection (TLS1) detuned by Δλ1=−70pm. A first input optical pulse of duration w (red line) triggers a partial or complete spike response (blue line). A second pulse, delayed by a time s (also red), tests whether the MRR remains in a refractory state. Different pulse durations w ranging from 64 to 160 ns (rows), and test-pulse delays s ranging from 32 to 384 ns (columns), are explored. The test pulse duration is fixed at 48 ns in all scenarios. All optical pulses have an optical power of 12 mW.

Fig. 9. MRRs’ spiking patterns obtained for input sequences of optical pulse with diverse durations: several μs (left), 200 ns (center), and 50 ns (right), interleaved by 200 ns in all cases. The detuning operation is set to [Δλ1,Δλ2]=[80,80]pm. Yellow, green, and purple lines indicate the configurations where λ1 only (yellow), λ2 only (green), or both λ1 and λ2 channels (purple) are open. Refractoriness is not observed for this wavelength detuning.