A photonic synapse at 239 picojoules narrows the one gap tissue was supposed to own

What the work claims

This is a device physics paper with neuromorphic demonstrations. Sahu, Ghosh, Ghoshal, Das and Ray claim three things. First, a microcavity chemical vapor deposition process grows uniform, large-area mono- and few-layer Mo0.5W0.5S2 alloy films, solving a known growth problem for 2D alloys. Second, the resulting optoelectronic synaptic device, driven by growth-induced sulfur vacancies, reproduces a checklist of biological synaptic behaviors: excitatory postsynaptic currents, paired-pulse facilitation of about 170 percent, stimulus-dependent short- and long-term plasticity and memory, learning-forgetting-relearning, and Pavlovian associative learning using two different optical wavelengths. Third, the measured conductance weight updates fed into a simulated 784-256-10 network reach 92.43 percent accuracy on MNIST, with retention degradation quantified over time.¹

The headline energy figure: a minimum of 239 pJ per single optical synaptic training event, computed as spike current times read voltage times pulse duration, for a 500 ms pulse at 0.5 V read bias. The authors state this is about four times lower than a silicon CMOS optoelectronic synaptic device they cite, and they place it against the biological synapse's 1 to 100 fJ per event with the explicit note that shorter pulses and lower read voltages should close the gap further.¹

How it works

The material is the mechanism. Alloying molybdenum and tungsten disulfide in a microcavity reactor confines precursor vapors and yields uniform large-area films, and the growth process leaves intrinsic sulfur vacancies that X-ray photoelectron spectroscopy, Kelvin probe force microscopy and scanning transmission electron microscopy all confirm. Those vacancies are trap states: 405 nm optical pulses generate photocarriers, and the competition between trapping, detrapping and recombination produces the volatile conductance transients that read as short-term plasticity, while longer or repeated stimulation moves the device into persistent, long-term conductance change.¹

The characterization is careful where it counts. The paired-pulse facilitation index peaks at 170 percent with 100 ms between two consecutive pulses, against 121 percent for MoS2 and 118 percent for WS2 control devices, a comparison that attributes the enhancement to alloying rather than to device geometry. Plasticity is stimulus-dependent in the biological sense: pulse width, pulse count, frequency, optical power and read bias all reshape the response, and the device cycles between short-term and long-term regimes on demand. The Pavlov demonstration exploits wavelength selectivity: a 375 nm pulse plays the role of the unconditioned stimulus and a weak 405 nm pulse the conditioned one, and after pairing, the 405 nm pulse alone evokes the amplified response.¹

The systems-level claim is a simulation, stated as such. Measured conductance weight-update curves, with their nonlinearity and retention drift, are fed into a 784-256-10 ReLU network trained on MNIST, reaching 92.43 percent recognition accuracy, with the paper quantifying how accuracy decays as retention is degraded. A separate demonstration uses the wavelength-selective photoresponse to perform color image filtering, extracting a single color channel from a mixed image.¹

Where a skeptic should push

The energy number needs context before it travels. The 239 pJ figure is a minimum, taken at the gentlest operating point the device supports: a 500 ms optical pulse and a 0.5 V read bias, with energy scaling linearly in both pulse duration and read voltage. Speed and energy are being traded against each other, and the paper's own framing, that shorter pulses and lower biases will reduce energy further, is a program statement, not a measurement. For always-on temporal processing, where events arrive at kilohertz rates rather than over half-second windows, the device as characterized would be far slower than the workloads neuromorphic silicon already runs.¹

Second, everything systems-level is simulated. The MNIST result demonstrates that the measured weight-update characteristics are good enough to encode a trained network; it does not demonstrate learning on the device, an array of devices, or any crossbar integration, and the retention analysis shows the accuracy is sensitive to how well weights hold. Third, the comparisons are against selected references, as is normal in device papers: the four-fold-better-than-silicon claim is relative to one cited CMOS optoelectronic synapse, not to the best published memristive or ferroelectric devices, some of which already report femtojoule-class switching. And the distance to biology remains large: 239 pJ is roughly four to six orders of magnitude above the 1 to 100 fJ the paper itself cites for a biological synaptic event.¹

Demonstrated: a reproducible route to uniform 2D alloy films and a single device that emulates the phenomenology of synaptic plasticity at documented energies. Asserted: that this architecture is a promising path toward the efficiency of neural processing. Unmeasured: array-level behavior, write endurance, device-to-device variability across a wafer, and any end-to-end energy accounting that includes the laser or LED drive electronics.

What pJ-class photonic synapses change for wetware's energy case

For organoid intelligence, the device matters less as competition than as a measuring stick the field should internalize. The unique selling proposition of biological computing has been that no engineered substrate comes within orders of magnitude of the synapse's energy budget. Papers like this one compress that claim on both sides at once: engineered synapses keep sliding down the energy axis while carrying more and more of the plasticity phenomenology, paired-pulse facilitation, short- and long-term memory, associative conditioning, that used to be cited as the tissue's exclusive repertoire. At 239 pJ today with an explicit roadmap to femtojoules, the window in which living tissue can rest on an energy argument is narrowing on a predictable curve.¹

The non-obvious implication runs in the other direction, and it is about interface geometry, not efficiency. This device is written with light and read electrically, at picojoule-class energies, with plasticity that depends on wavelength. That is precisely the write-read asymmetry organoid interfaces want: optical stimulation is already the least invasive, most spatially selective way to perturb neural tissue, optogenetics being the proof, and this paper shows the same physical channel can carry plasticity-specific signals. The Pavlov demonstration is the transferable idea in miniature: two wavelengths, one strong and one weak, paired to induce association, executed entirely in the optical domain. An organoid photostimulation rig that assigns different wavelengths to different plasticity timescales, fast modulatory versus slow consolidating signals say, would be borrowing a control scheme that this device demonstrates in a passive film.¹

The threat is definitional. If artificial synapses keep absorbing the phenomenology, the remaining case for computation on living tissue concentrates into what a vacancy-trap film cannot do: self-organization, developmental scaling, homeostatic self-repair, biochemical signaling, and learning that rearranges the hardware rather than merely its conductance states. That is still a real case, but it is a harder, longer-term case than energy, and it fails if the field keeps letting energy dominate its public arithmetic. The honest benchmark this paper hands over is its own: cite the biological number, 1 to 100 fJ per event, and report how far your substrate is from it, with the operating point stated. An organoid computing claim that cannot state its joules-per-effective-synaptic-event, support included, is not ready to share a chart with this device.¹

The bottom line

Established: a microcavity CVD route to large-area MoWS2 alloy films, and a single optoelectronic synaptic device that reproduces excitatory postsynaptic currents, paired-pulse facilitation near 170 percent, short- and long-term plasticity and Pavlovian associative conditioning at a documented minimum of 239 pJ per optical event, with measured weight updates good for 92.43 percent MNIST accuracy in simulation. Unestablished: array-level learning, speed at neuromorphic rates, femtojoule operation, and any end-to-end energy claim including drive electronics. What would confirm the trajectory: sub-100 fJ operation at microsecond-class pulse widths on an integrated array. What organoid intelligence should take from it now: the energy gap to biology is a moving target engineered substrates are closing methodically, so the field's case must rest on capabilities no trap-state film will acquire, and its photostimulation interfaces should be stealing this paper's wavelength-multiplexed plasticity control.