Research analysis · Devices and interfaces

A van der Waals floating-gate synapse that learns through light

A single ReS2/hBN/InSe transistor from Dongguk University combines nonvolatile charge-storage memory with broadband optical weight updates across 532 to 850 nm. It reproduces paired-pulse facilitation, Pavlovian associative and discrimination conditioning, and reconfigurable Boolean logic in one device, with an on/off ratio above 10^6 and retention near 10^4 seconds.

Source: A van der Waals Floating-Gate Platform for Broadband Optoelectronic Neuromorphic Computing, arXiv:2610.34738, 2026. Primary source. Read the full 24-page preprint PDF, including methods and figure captions.

What the work claims

This is a primary device result: one carefully engineered transistor that does the two things a neuromorphic synapse must do, store a weight nonvolatilely and update it with low-energy stimuli, and adds a third capability most artificial synapses lack, updating by light over a broad spectrum. The stack is a van der Waals heterostructure: an ReS2 channel that is photoresponsive, a hexagonal boron nitride (hBN) layer about 10 nm thick that acts as the tunneling dielectric, and an InSe layer that acts as the floating gate where charge is stored.1

The claimed numbers are strong for a single device. The memory window, the threshold-voltage separation between programmed and erased states, reaches about 93 V under a plus/minus 60 V gate sweep, with an extracted stored charge density of 6.67 x 10^12 cm^-2. The conductance on/off ratio exceeds 10^6. The device survives 1,000 program/erase cycles and holds both states for roughly 10^4 seconds without refresh. On the synaptic side, the group demonstrates short- and long-term plasticity, pulse-amplitude- and pulse-duration-dependent plasticity, paired-pulse facilitation whose index decays with inter-pulse interval, and bidirectional weight control in which 532 nm light potentiates and electrical pulses depress. Beyond single-synapse behavior, they report Pavlovian associative conditioning, a harder discrimination-conditioning protocol with two conditioned stimuli, and Boolean AND and OR logic reconfigured by changing stimulation conditions.1

How it works

The physics is classical floating-gate memory rebuilt in two-dimensional materials. A positive back-gate pulse drives electrons from the ReS2 channel across the hBN tunneling barrier into the electrically isolated InSe layer; the trapped electrons deplete the channel and create the low-current program state. A negative pulse reverses the field, ejects the stored electrons, and restores the high-current erase state. Control transistors fabricated without the InSe layer show negligible hysteresis, which the authors use to attribute the memory window to charge storage in the floating gate rather than to contact or channel artifacts. The memory window grows with the sweep range, reaching its maximum among tested devices at an hBN thickness near 10 nm, the expected competition between tunneling probability during writing and leakage during retention.1

The optical pathway is what makes the device interesting as a synapse rather than as memory. ReS2 absorbs across the visible and near-infrared, so illumination generates photocarriers in the channel that modulate conductance without any gate pulse. At a fixed power density of 1.7 mW/cm^2, distinct postsynaptic responses are elicited at 532, 650, and 850 nm, with the strongest response at 532 nm. Consecutive 532 nm pulses (2.5 mW/cm^2, 0.2 s) produce cumulative potentiation, while negative gate pulses produce depression, giving bidirectional weight control with two physically distinct write channels, one optical and one electrical.

The higher-order demonstrations follow directly from that plasticity. In the conditioning experiment, 15 V gate pulses serve as one conditioned stimulus and minus 15 V pulses as another, while a 532 nm optical pulse (0.5 mW/cm^2, 0.2 s) serves as the unconditioned stimulus, with a 1 nA postsynaptic current defined as the salivation-like response threshold. After association training, the reinforced stimulus alone crosses threshold while the unreinforced one does not; the discrimination version trains the device to respond to one sign of gate pulse and not the other. The logic gates exploit the same dual-input structure: a minus 40 V electrical pulse and a 532 nm optical pulse either combine (AND) or either suffices (OR), depending on the bias conditions.

Where a skeptic should push

The most load-bearing number is also the most quietly alarming one. A 93 V memory window is only reachable with plus/minus 60 V sweeps; even the synaptic demonstrations run at tens of volts, and the program/erase pulses are plus/minus 60 V. That is one to two orders of magnitude above the supply voltages of any CMOS-compatible peripheral, so every energy-efficiency argument about the device currently sits next to a driver circuit that would dominate the energy budget. The authors do not report an energy-per-update figure in the main text, which is exactly the number a neuromorphic claim needs.

Second, the durability metrics are bench-scale. Retention of 10^4 seconds is about 2.8 hours, fine for a laboratory demonstration but far from the years expected of nonvolatile memory, and 1,000 program/erase cycles is modest; production nonvolatile devices tolerate 10^5 or more. Third, the learning demonstrations are device-level, not system-level. Pavlovian conditioning in a single transistor with a manually set 1 nA threshold is a faithful emulation of one synapse, not a trained network; no classification task on a standard dataset is shown. Fourth, everything is measured at room temperature on individual devices, with no crossbar array, no write disturb data at scale, and no inter-device variability statistics. The results are exactly what the authors claim and no more: an impressive multifunctional device, not yet a technology.

Optical weight writing and the organoid interface

The non-obvious implication for organoid intelligence is architectural. Every serious closed-loop program for biological computing faces the same I/O asymmetry: you read living neural tissue electrically, through MEA electrodes, but the most selective way to actuate it is optical, through optogenetically expressed actuators. This device is the mirror image of that asymmetry, electrical read and optical write, demonstrated in a single two-terminal-gated element with a retention time that exceeds any experimental session. That makes it a plausible building block for the peripheral layer of an organoid training rig: an optoelectronic weight store that buffers stimulation patterns, holds them nonvolatilely for the duration of a culture session, and releases them as patterned light while the culture's electrical answer is read out on electrodes.

There is a sharper research angle. The device's write pathway works from 532 to 850 nm, and the near-infrared end of that band overlaps the window where biological tissue is most transparent. A 650 to 850 nm optical write channel is precisely the range in which patterned optogenetic stimulation of a cortical organoid avoids the scattering and absorption that defeat shorter wavelengths in thick tissue. The authors built this for silicon neuromorphics, but the specification, broadband optical potentiation at sub-mW/cm^2 power densities with electrical depression for balance, reads like a wish list for the photostimulation side of a biological computer, and the paired-pulse facilitation data give concrete timing parameters (sub-second intervals) for what an optical training pulse train would need to look like.

The threat is just as concrete. This device shows that in-memory learning with optical I/O does not require living tissue at all, and its spec sheet defines what the biological substrate must beat: a weight that holds for hours, updates bidirectionally in 0.2 s pulses, survives 1,000 updates without degradation, and runs on a 20 W-class system budget rather than an incubator plus perfusion plus antibiotics. A cortical organoid's advantage is not any single synaptic metric but massive parallelism, adaptive self-repair, and computation in the same element that stores the state. If the field cannot demonstrate a task where living tissue outperforms a stack of devices like this one on latency, energy, or training data efficiency, the economic argument for wetware narrows to niches. That comparison has rarely been made honestly, because the device and the tissue are usually benchmarked on different tasks. This paper makes the comparison easier to force, which is a service to both sides.

The bottom line

Established: a ReS2/hBN/InSe floating-gate transistor stores charge nonvolatilely with a 93 V memory window, an on/off ratio above 10^6, 1,000-cycle endurance, and 10^4-second retention, and the same device supports broadband optical potentiation, electrical depression, paired-pulse facilitation, associative and discrimination conditioning, and reconfigurable Boolean logic. Not established: system-level learning performance, CMOS-compatible operating voltages, array-scale behavior, or a competitive energy-per-update figure. What would confirm the platform: crossbar integration with write-disturb and variability statistics, an energy-per-update measurement, and a drop in operating voltage toward single digits. What would weaken it: retention or endurance that fails to scale, or optical write energies that exceed the electrical ones by orders of magnitude. For organoid intelligence, the durable residue is a spec sheet and an architecture: the optical-write/electrical-read device is the peripheral design pattern that closed-loop biological computing will need, and simultaneously the baseline competitor that defines what living synapses must demonstrably beat.

Frequently asked questions

What is a floating-gate synaptic transistor?

A transistor with an electrically isolated charge-storage layer between its gate and channel. Charge injected into that layer shifts the channel's threshold voltage persistently, so the conductance encodes a weight that survives without power, and pulses can move the weight up or down the way synaptic plasticity moves a biological synapse's strength.

Why does the optical write channel matter?

It decouples writing from the gate voltage entirely: light at 532 to 850 nm generates carriers in the ReS2 channel that potentiate the synapse, while electrical pulses depress it. Two physically independent write paths give bidirectional weight control and open the near-infrared band, where light penetrates tissue most effectively.

What does the Pavlovian demonstration actually show?

That a single device can bind two stimuli into an association. Gate pulses of one sign are paired with an optical pulse until the gate pulse alone pushes the postsynaptic current past a 1 nA threshold; the discrimination version trains the device to respond to one sign and not the other. It demonstrates history-dependent plasticity in one synapse, not network learning.

What are the main engineering weaknesses?

Operating voltages of tens of volts, with program/erase at plus/minus 60 V; retention of about 2.8 hours rather than years; 1,000 program/erase cycles; and no reported energy-per-update figure. All are single-device, room-temperature results without array integration or variability statistics.

Why should an organoid-intelligence program care about a silicon device?

Because it defines the competitor and the interface. The optical-write/electrical-read architecture is exactly the peripheral pattern a closed-loop organoid rig needs, and its published specifications (retention, update energy per area, pulse timing, endurance) are the honest yardstick against which the claimed advantages of living synapses, parallelism, self-repair, and in-memory computation, should be measured.

References

  1. T. P. A. Bach and S. Cho. A van der Waals Floating-Gate Platform for Broadband Optoelectronic Neuromorphic Computing. arXiv:2610.34738. 2026. https://arxiv.org/abs/2610.34738. Accessed 2026-10-31.