A two-terminal photosynapse that remembers the light it saw

What the work claims

Sahu, Ghosh, Ghoshal, Das, and Ray report a primary experimental device study from IIT Kharagpur: a gate-free, metal-semiconductor-metal optoelectronic synaptic device (OSD) using a ternary MoWS₂ alloy as the only light-sensitive material, grown by a space-confined microcavity CVD reactor and characterized with Raman, photoluminescence, synchrotron XRD, XPS, KPFM, and electron microscopy¹. The claim is that growth-induced sulfur vacancies make the film intrinsically synaptic, emulating excitatory postsynaptic current (EPSC), paired-pulse facilitation, short- and long-term plasticity, learning-forgetting-relearning, and Pavlovian associative learning, with measured conductance curves fed into an MNIST network simulation (92.43% accuracy) and a hardware-level color-filtering demonstration.

How it works

The mechanism is defect-mediated persistent photoconductivity. The microcavity reactor holds sulfur chemical potential quasi-constant over the growth zone, producing uniform large-area mono- and few-layer alloy films; XPS gives a stoichiometry of 0.43:0.47:1.78 for Mo:W:S, clearly sulfur-deficient, and KPFM maps a highly non-uniform surface potential that the authors attribute to electron and hole trap sites¹. Sulfur vacancies introduce in-gap states. Photons with energy above the bandgap generate electron-hole pairs; a fraction of the electrons is captured by vacancy traps instead of recombining, and thermal release back to the conduction band is slow. The result is a photocurrent that outlives the pulse: a conductance that encodes the recent optical history.

The synaptic phenomenology follows directly. A pair of 405 nm pulses 100 ms apart gives a maximum paired-pulse facilitation of about 170% (second EPSC over first), versus 121% for a MoS₂ control and 118% for WS₂; the facilitation decays toward about 150% as the interval stretches to 8 s, with a double-exponential fit yielding a fast time constant of 1.188 s and a slow one of 223 s. After 80 s of illumination the EPSC decays with time constants of 352 s (trap-limited) and 35.6 s (band-to-band), nearly an order of magnitude slower than the binary controls. Conductance potentiation grows sub-linearly with pulse count (nonlinearity beyond about 250 pulses) and with optical power between 1.68 and 11.21 mW/cm². The device also shows spike-frequency-dependent plasticity with a fitted cutoff near 0.95 Hz, emulating the high-pass behavior of a biological synapse, and wavelength-dependent strength: strong synaptic response at 350, 375, and 405 nm, weak beyond 405 nm, enabling a two-color Pavlov conditioning demo (375 nm as unconditioned, 405 nm as conditioned stimulus, threshold 16.36 nA)¹.

The systems numbers are honest about their limits. Energy per synaptic event is 239 pJ at a 0.5 V read bias and 500 ms pulse, which the authors state is about four times lower than a silicon CMOS optoelectronic synapse, but still roughly three to five orders of magnitude above a biological synapse's 1 to 100 fJ. The MNIST result is a simulation: a 784-256-10 network whose first-layer weights are quantized to the measured conductance levels (27.97 to 73.01 nS, fitted nonlinearity factor 0.738), reaching 92.43% shortly after programming. The most interesting system result is the drift study: applying the device's measured Kohlrausch decay to the mapped weights collapses classification accuracy to 8.92%, essentially chance, by 10,000 s, with the confusion matrix degrading into a single dominant output class¹. The conductance memory decays on the timescale of a long experiment, and the network's apparent robustness holds only so long as the decay stays quasi-uniform.

Where a skeptic should push

The load-bearing assumption is that sulfur vacancies are both the necessary and the intended mechanism. The evidence is strong but circumstantial: correlated XPS sulfur deficiency, KPFM potential heterogeneity, TEM-visible vacancies, slow decay constants, and sub-linear EPSC-versus-power scaling all point at trap-mediated persistent photoconductivity. But trap spectroscopy is absent; the deep versus shallow level assignment is inferred from decay fits, which are famously non-unique. Nor is there any cycling endurance data, device-to-device statistics, or environmental stability (oxygen, water, temperature), all of which decide whether a vacancy-rich 2D film survives outside a glovebox.

Second, the neuromorphic framing outruns the demonstration in places. The Pavlov emulation is a threshold-crossing argument, not associative storage: conditioning works because two wavelengths sum past a preset current. The MNIST network is trained in software and then projected onto measured conductance levels; the device does not learn. The color filtering is a calibrated response-function mapping, not in-sensor computation. And the operating points are modest: 500 ms pulses, 10 mW/cm² illumination, micron-scale channels, nanoamp signals. This is a materials-and-device paper wearing a computing costume, and it is better read that way.

What a photosynapse means for the optical MEA channel

Arrays that combine electrical recording with optical stimulation already juggle a metrology problem: the light used to perturb the tissue (or to excite indicators) is rarely measured where it lands. The device's properties sketch a per-pixel optical witness element co-integrated with the electrode. It responds in the blue-UV band where channelrhodopsin-family stimulation and many voltage-indicator excitation lines live, it consumes nanoamps of dark current, and its conductance is not just proportional to dose but encodes the recent dose history, with an adapting high-pass response whose cutoff sits near 1 Hz. A pixel like that would report delivered photon flux as a current while automatically suppressing the constant background and emphasizing transients: exactly the preprocessing an optical stimulation monitor needs, and it does the temporal filtering in the device rather than in firmware. The wavelength selectivity adds a second channel of information, since 405, 525, and 630 nm are potentiated differently, letting a single element separate stimulation light from imaging light by conductance signature rather than by filters and extra detectors.

The threat is the same defect physics read on a longer clock. The vacancy traps that produce 170% facilitation and 352 s retention also make the device's transduction gain a slowly decaying function of everything it has seen. The paper's own drift study is the cautionary tale: within 10,000 s, roughly the length of a long organoid recording session, mapped conductance states relax enough that a downstream classifier built on them collapses to chance. Any calibration curve of the form "this current means that irradiance" would quietly become wrong over exactly the timescale on which an MEA experiment runs, and because the decay is history-dependent (it depends on how many pulses potentiated the device), it cannot be corrected by a single lookup table. An optical monitor, an on-chip calibration reference, or an analog trim built from this materials family would need periodic re-referencing against a known optical standard during the recording itself.

The non-obvious point is that the opportunity and the threat cannot be separated by engineering the traps away, because they are the same trap spectrum. Strong memory, low dark current, retinal adaptation, and slow drift are one physical object. The instrumentation question this paper leaves on the table is which side of that trade a given pixel should sit on: a fast, memoryless optical dosimeter requires suppressing exactly the vacancy density that makes this device interesting. That is a real design tension for anyone planning to put photosensitive 2D materials on an active array substrate, and it is sharper than the paper's own neuromorphic framing suggests.

The bottom line

Established by measurement: a CVD ternary MoWS₂ alloy film is intrinsically synaptic under optical stimulation, with 170% paired-pulse facilitation, multi-hundred-second retention, wavelength-selective plasticity, and 239 pJ per event. Established by simulation, not by the device: MNIST recognition and robustness claims. The drift data, taken at face value, is the most instrumentation-relevant result in the paper: history-dependent conductance relaxation collapses a mapped network within hours. For MEA hardware, the near-term significance is a design template for adaptive optical witness pixels alongside recording electrodes, paired with a standing warning that defect-engineered photosensors carry their own slowly fading memory into any calibration chain.