Light-Tuned Artificial Synapse Made From Doped WSe2 Mimics the Brain and Runs Neural Networks

Researchers in South Korea, Pakistan, Belgium, China, Saudi Arabia and the United States have unveiled an optically driven artificial synapse built from chemically doped two-dimensional tungsten diselenide, a device that can sense light, remember what it has seen, perform logic operations and even serve as the building block of a neural network that recognizes handwritten digits. The work, published in Advanced Composites and Hybrid Materials, addresses one of the most pressing bottlenecks in modern computing: the enormous energy cost of shuttling data between memory and processors in conventional von Neumann architectures. By merging sensing, memory and computation into a single nanoscale component, the team led by Ghulam Dastgeer and Jonghwa Eom of Sejong University offers a glimpse of hardware that computes the way biology does, with signals that strengthen or weaken depending on how often and how intensely they arrive.

The heart of the device is a tri-layer stack of tungsten diselenide, a semiconductor only a few atoms thick that belongs to the family of transition metal dichalcogenides. Two-dimensional materials like WSe2 are prized in neuromorphic engineering because their entire volume sits at the surface, making their electronic properties exquisitely sensitive to light, electric fields and chemical modification. In their native form, however, WSe2 devices often fall short of the stable, tunable conductance states that artificial synapses demand. The researchers solved this by treating the material with benzyl viologen, a strong electron-donating molecule that donates charge to the WSe2 lattice and shifts its Fermi level, effectively engineering the carrier polarity of the channel and setting the stage for the dramatic optical behaviors that follow.

Before any synaptic claims could be made, the team had to prove the material survived the doping process intact. Atomic force microscopy confirmed the structural integrity and surface morphology of the tri-layer films, while Raman spectroscopy verified that the crystal lattice remained high quality after exposure to the benzyl viologen treatment. On the electrical side, gate-dependent transfer curves and current-voltage measurements mapped out how the doping transformed the device’s behavior, and the researchers constructed an energy band diagram showing the Fermi level shift that the molecular dopants induce. This combination of structural and electrical characterization matters because any degradation of the delicate 2D lattice would undermine the reproducibility that real neuromorphic circuits require.

The most striking results emerge when light hits the device. The BV-doped WSe2 synapse responds strongly to both visible and deep-ultraviolet illumination, but it is the deep-ultraviolet response that gives the device its memory. Under DUV light, charges become trapped within the material system, and these trapped charges persist long after the light is switched off, producing a stable conductance state that functions as a nonvolatile memory. In other words, a flash of ultraviolet light writes information into the device, and the device holds that information without power, much as a biological synapse retains the trace of a past experience. This optically programmable memory is what allows the same physical structure to act simultaneously as a photodetector and a synaptic element.

When stimulated with sequences of optical pulses rather than continuous illumination, the device reproduces hallmark behaviors of biological synapses. Paired-pulse facilitation, a form of short-term plasticity in which the second of two closely spaced signals produces a larger response than the first, appears naturally in the device’s conductance dynamics. With longer or more intense stimulation, the device exhibits long-term potentiation and long-term depression, the persistent strengthening and weakening of connection strength that neuroscientists consider the cellular basis of learning and memory. Crucially, the magnitude of the response is tunable: by controlling the duration of the deep-ultraviolet pulses, the researchers increased the device’s photoresponse from 25.2 percent to 1438.7 percent, a dynamic range wide enough to encode many distinct synaptic weights.

That tunability translates directly into computational capability. The team demonstrated that the device can perform a logic OR gate operation, outputting a high signal when either of two inputs is present, and that it can carry out selective alphabet recognition, distinguishing among written characters based on the optical signatures stored in its conductance states. These demonstrations show that the device is not merely a passive memory element but an active processing unit, capable of the kind of in-memory logic that could dramatically reduce the energy overhead of edge-computing systems, where data must be processed locally under strict power budgets rather than shipped to energy-hungry data centers.

To test whether the device’s synaptic behavior could scale to meaningful machine learning, the researchers implemented its experimentally measured synaptic weight characteristics in an artificial neural network and trained that network to identify handwritten digits from the MNIST dataset, a standard benchmark in computer vision. The simulated network achieved an accuracy of 92 percent, and the team validated the learning process through epoch-dependent learning curves and confusion matrix analysis, confirming that the network’s errors were distributed in the way one expects from a genuinely learning system rather than a lucky configuration. The result is significant because it grounds the device physics in a concrete application: the same conductance modulation that the researchers measured in the laboratory can, in principle, serve as the adjustable weights that artificial neural networks rely on.

The broader significance of the work lies in its multifunctionality. Most neuromorphic devices reported to date specialize in one role, whether as memristors, phase-change memories or electrochemical transistors, and require separate photonic or electronic circuitry to interface with the outside world. By combining optical sensing, nonvolatile charge-trapping memory, short- and long-term plasticity, logic operation and pattern recognition in a single doped 2D material platform, the WSe2 synapse points toward integrated intelligent sensors that could see, remember and compute in one package. Such devices would be natural fits for autonomous drones, biomedical implants and environmental monitors, where every milliwatt saved extends operational life.

There are, of course, hurdles between a laboratory demonstration and deployable technology. The device relies on exquisitely controlled chemical doping of atomically thin crystals, and scaling such processes to wafer-scale manufacturing while maintaining the uniformity that large neural networks demand remains an open challenge in the 2D materials community. The 92 percent MNIST accuracy, while respectable, still trails the best software-based networks, suggesting that device variability and nonlinearity in the conductance updates will need further engineering. The published version of the paper is also an early-release, peer-reviewed accepted manuscript that may undergo further edits before the final version of record appears, so specific figures could be refined.

Even so, the study adds an important entry to the growing catalog of optoelectronic synapses, and its use of benzyl viologen doping to engineer carrier polarity in tri-layer WSe2 gives other researchers a concrete chemical lever to pull. The work was supported by the Korea Basic Science Institute through a National Research Facilities and Equipment Center grant funded by the Korean Ministry of Education, by an IITP Information Technology Research Center grant from the Ministry of Science and ICT, and by the ongoing research funding program of King Saud University in Riyadh. As artificial intelligence migrates from the cloud toward the edge, devices that can be programmed by nothing more than pulses of light, and that learn the way neurons do, may prove to be exactly the kind of hardware the next generation of intelligent machines is waiting for.

Subject of Research: Optoelectronic synaptic devices based on doped two-dimensional tungsten diselenide for neuromorphic computing and logic operations

Article Title: A multifunctional optoelectronic synaptic device for logic operation and neuromorphic computing

Article References: Dastgeer, G., Nisar, S., Fatima, I., Ali, R. F., Rabani, I., Hussain, K., Alsalme, A., & Eom, J. (2026). A multifunctional optoelectronic synaptic device for logic operation and neuromorphic computing. Advanced Composites and Hybrid Materials. https://doi.org/10.1007/s42114-026-02103-z

Image Credits: AI Generated

DOI: 10.1007/s42114-026-02103-z

Keywords: neuromorphic computing, optoelectronic synapse, tungsten diselenide, WSe2, benzyl viologen doping, two-dimensional materials, memristor, synaptic plasticity, charge trapping, artificial neural network, MNIST, logic gates