One laser, two neuron types: photonic spiking neurons on silicon

What the work claims

This is a primary experimental device paper from the University of West Attica, CEA-Leti, and Scintil Photonics. The authors fabricate a photonic integrated circuit carrying an array of two-section gain and saturable-absorber (SA) lasers built from InP quantum wells, monolithically integrated on a silicon substrate, and they map the self-pulsation dynamics of these devices in detail.¹

The central claim has two parts. First, by tuning only the DC gain current and the reverse bias on the absorber section, the same physical laser exhibits two distinct excitable regimes that map onto two textbook neuron types: a resonate-and-fire-like regime (Regime I), in which pulse repetition frequency stays near 16 MHz almost regardless of drive strength while spike amplitude grows with current, and an integrate-and-fire-like regime (Regime III), in which repetition frequency climbs strongly with input up to 1.8 GHz while amplitude saturates. Both behaviors live in one device, selected by a minor bias change. Second, the team validates the integrate-and-fire isomorphism directly, demonstrating a firing threshold, sub-threshold temporal summation, and a finite refractory period, and they map how cavity length and the gain-to-absorber length ratio control where the fast regime exists.

How it works

A two-section laser is a cavity split into a pumped gain section and an unpumped absorber section. In Regime III the physics is Q-switching: carriers accumulate in the gain section until the rising intracavity intensity bleaches the saturable absorber, the cavity dumps its stored energy in one short optical pulse, absorption recovers, and the cycle restarts. Harder input drives faster carrier accumulation, so spike rate encodes input strength. That is the photonic analog of leaky integrate-and-fire: integrate until threshold, fire, reset, repeat. In Regime I the mechanism is different: the pulses are attributed to canard-type spikes driven by optothermal dynamics, giving an almost input-independent repetition rate with amplitude carrying the signal, which is the signature of resonate-and-fire behavior.¹

The neuronal validation is the most interesting experiment in the paper. Biased near the edge of Regime III (71.24 mA gain current, -2.2 V absorber bias), the laser is probed with 0.5 ns electrical pulses. A single 0.8 V pulse reliably produces one optical spike; a 0.3 V pulse does nothing. Two 0.3 V pulses spaced 0.4 ns apart summate to trigger a spike, but the same pair spaced 1.1 ns apart does not, showing that the accumulated excitation decays on a sub-nanosecond timescale, a membrane-time-constant analog. Two suprathreshold 0.8 V pulses 0.5 ns apart yield only one spike, while at 0.8 ns spacing they yield two, bounding the refractory period between those values; the authors note the device's maximum repetition frequency implies a minimum resting period of about 0.55 ns.¹

The design mapping is equally concrete. Across nine laser variants (cavity lengths of 400, 600, and 800 micrometers; absorber length ratios of 3, 5, and 8 percent), devices with 3 or 5 percent absorber ratios show all three dynamical regimes, while ratios above 8 percent collapse the device to the slow resonate-like regime alone, because a longer absorber stores more saturation energy and demands more intracavity energy to bleach. Longer cavities widen the bias window for fast spiking and push the maximum repetition frequency higher. One counterintuitive result: the highest repetition frequencies occur at a 5 percent absorber ratio rather than the lowest, 3 percent, because the fast regime switches on at lower absorber bias there, reducing effective intracavity loss. For anyone designing fast integrate-and-fire laser neurons, the authors' rule of thumb is that the absorber ratio should not exceed 5 percent.¹

Where a skeptic should push

The load-bearing assumption is that dynamical analogy at the single-device level implies a path to useful neuromorphic systems. Everything demonstrated here is one laser at a time: no interconnected network, no silicon waveguide synapses doing weighted routing (they are explicitly future work), no learning, no task. The "neural isomorphism" is a claim about three response properties of an excitable system, threshold, summation with a leaky memory, and refractoriness, which many nonlinear oscillators exhibit; the bar for calling something a neuron in a computing sense is what it does in a network.

Several specifics deserve weighting. The device's external quantum efficiency is modest (24.6 microwatts per milliampere at zero absorber bias, falling to 13.1 at 3 V), which the authors attribute to the low reflectivity of their Sagnac-loop cavity facets; energy per spike and wall-plug efficiency, the numbers that decide whether photonic neurons beat electronic ones, are not reported. The data underlying the results are not publicly available. The manuscript still carries unreplaced journal submission placeholders ("Received XX Month XXXX"), so treat it as a preprint of work likely headed to an Optica-family journal, not as peer-reviewed record. And the refractory bound deserves a note of interpretive care: the 0.5 to 0.8 ns two-pulse window demonstrates recovery of excitability but was probed at a single bias point, so it should not be quoted as a universal device spec.¹

Speed is not where living tissue competes

For organoid intelligence, the uncomfortable arithmetic is the refractory period. This laser recovers in roughly half a nanosecond and spikes at gigahertz rates; a cortical neuron refracts for one to two milliseconds. That is a six-order-of-magnitude gap in raw spike throughput, and no plausibly engineered culture closes it, because the timescale is set by membrane biophysics, not by incremental improvements in stem-cell protocols. The corollary: any workload framed as "spike this fast" is permanently ceded to photonics and electronics. The honest pitch for biological computing has to be about what slow tissue does that fast lasers demonstrably do not, namely self-modify its wiring under activity and repair its own components, and the field should be benchmarking those properties rather than spike rates.

The non-obvious implication cuts the other way as well, and it is methodological. The paper's most transferable result is not the device but the demonstration that the computational regime of a single physical element is a control variable, not a fabrication accident. One bias knob switches the element between an integrator and a resonator; the authors prove this by mapping a regime diagram over nine device variants. A cultured neural network has the same property at a different timescale: whether a dish computes in a bursting or a tonic regime is a dynamical state, selectable in principle through stimulation and neuromodulation, and it is rarely treated that way. Organoid labs tend to characterize whatever regime the culture falls into; this paper is a template for treating regime selection as the first-class experimental object, with systematic parameter maps instead of single-condition showcases.

The hybrid opportunity is real but bounded. A photonic layer offers microsecond-latency transduction and dense wavelength-multiplexed routing that no multielectrode array matches, so photonics is a plausible periphery for stimulation and readout around living tissue. But note the mismatch the paper makes vivid: the photonic neuron operates on a half-nanosecond refractory cycle while tissue speaks in milliseconds, so a photonic front-end wired to a culture will idle between tissue events. The rational architecture is photonics for transport and transduction, tissue for adaptation, with the interface deliberately clocked to the slow partner. The genuine threat is subtler than obsolescence: as silicon and photonic substrates absorb every well-posed spiking workload, the definition of "well-posed" migrates with them, and the window of tasks where living tissue is the natural answer narrows year by year unless the field produces regime-controlled, self-modifying demonstrations that fast hardware cannot mimic by bias tuning.

The bottom line

Established: one monolithically integrated III-V-on-silicon laser can be biased into either integrate-and-fire or resonate-and-fire behavior, the fast regime fires validated sub-nanosecond spikes up to 1.8 GHz, and cavity length plus absorber ratio give concrete design rules for keeping the fast regime open. Hypothesis: such laser neurons can be wired into scaled neuromorphic photonic circuits with useful learning; nothing in this paper tests that. What would confirm the systems claim is a multi-neuron PIC with weighted waveguide interconnects running a classification or reservoir task. What would weaken it is evidence that interconnection loss, thermal crosstalk, or per-spike energy at scale erases the speed advantage. For the biological-computing reader the takeaway is calibration, not alarm: photonics just claimed another six orders of magnitude of spike rate, which sharpens, rather than settles, the question of what slow living networks are for.