A single plasma treatment created the entire functional architecture.
At the intersection of materials science and neuroscience, researchers at Sungkyunkwan University have coaxed a thin crystal into behaving like a living nerve cell — sensing light and processing it in a single, unified act. By transforming rhenium selenide with a plasma treatment, they have built an artificial synapse that learns, remembers, and forgets much as biological tissue does. This is not merely a faster chip, but a rethinking of what computation can be: not calculation imposed on matter, but intelligence grown from its structure.
- Modern AI's hunger for visual data is straining conventional chips, which waste energy by separating sensing from processing — a split the biological brain never makes.
- Earlier attempts to use layered van der Waals crystals as artificial synapses kept failing due to uncontrollable grain boundaries, manufacturing residues, and inconsistent material quality across surfaces.
- A single plasma treatment of argon and hydrogen sulfide resolved these problems at once, converting only the top layer of rhenium selenide into nano-crystalline material while leaving the bulk crystal intact beneath.
- The resulting two-layer device replicates the architecture of a light-sensitive neuron, demonstrating long-term potentiation, memory transitions, and a 34.7% improvement in information retention over conventional bulk material.
- Tested against standard machine learning benchmarks, the device classified images with 96.24% accuracy — signaling that neuromorphic hardware capable of real-time visual processing at biological efficiency is within reach.
At Sungkyunkwan University, researchers have built an artificial synapse that responds to light the way a biological neuron does — not by simulating the process in software, but by engineering it directly into the structure of a crystal.
The material at the center of this work is rhenium selenide, a layered crystal so thin it exists at the atomic scale. The challenge it was meant to solve is fundamental: modern AI must process vast streams of visual data, but current chips handle sensing and processing as separate steps, burning energy and time in the gap between them. Biological synapses do both at once. Earlier attempts to replicate this using van der Waals crystals were undermined by manufacturing defects — uncontrollable grain boundaries, polymer residue, and warping that made consistent results impossible.
Professor Taesung Kim's team found their answer by studying the neuron itself. Light-sensitive ion channels embedded in a cell membrane control the flow of charged particles; the researchers recognized that a correctly engineered layered crystal could mirror this architecture. A single plasma treatment — argon and hydrogen sulfide applied to the surface of bulk rhenium selenide — converted only the topmost layer into nano-crystalline material, leaving the crystal beneath undamaged. The two layers together replicated the functional structure of a neuron: one mimicking the light-sensitive channel, one the cell's interior.
Scanning probe microscopy revealed how sulfur ions moved through the material, with grain boundaries in the nano-crystalline layer acting as precise gates. The device exhibited all the hallmarks of a real synapse — multiple conductance levels, long-term potentiation and depression, and transitions between short- and long-term memory — while retaining information 34.7% more efficiently than conventional bulk rhenium selenide.
In practical tests, the device performed edge detection on photographs and classified images from the CIFAR-10 dataset with 96.24% accuracy. Crucially, none of this required depositing extra layers or complex patterning — one plasma step created the entire functional architecture. The team sees this as a materials platform for neuromorphic semiconductors: chips that don't merely calculate, but process the world the way a brain does. Scaling and integration into real AI hardware systems is the work that comes next.
At Sungkyunkwan University, a team of researchers has engineered a device that learns the way a biological neuron does—by responding to light. The breakthrough centers on a material called rhenium selenide, a layered crystal so thin it exists at the atomic scale. By treating this material with a carefully designed plasma process, the team created something that mimics not just the function of a synapse, but its actual structure.
The problem they were solving is fundamental to the future of artificial intelligence. Modern AI systems need to process enormous amounts of visual information quickly—images, video feeds, sensor data streaming in from countless devices. Current computer chips handle this by separating the sensing and processing steps, which wastes energy and time. Biological brains don't work that way. A neuron's synapse—the junction where one nerve cell communicates with another—both senses light and processes that signal in a single integrated action. Researchers have long wanted to build artificial synapses that work the same way, using light-responsive materials. Layered van der Waals crystals seemed promising because they're naturally thin and respond well to light. But they came with stubborn technical problems: grain boundaries that couldn't be controlled, leftover polymer residue from manufacturing, warping at interfaces, and inconsistent quality across large areas.
Professor Taesung Kim's team approached the problem by looking at the actual architecture of a biological neuron. Light-sensitive ion channels sit in the cell membrane, controlling the flow of charged particles in and out of the cell. The researchers realized that layered van der Waals materials had a similar structure—if you engineered them correctly. They applied a single-step process using argon and hydrogen sulfide plasma to the surface of bulk rhenium selenide. This transformed just the top layer into nano-crystalline material made of tiny grains, while leaving the bulk crystal underneath intact and undamaged. The result was two integrated layers: one mimicking the light-sensitive ion channel, one mimicking the cell's interior.
Using scanning probe microscopy, the team traced how sulfur ions moved through the material. The grain boundaries in the nano-crystalline layer acted like gates, controlling exactly where and how the ions could travel. This gave them precise, deterministic control over how the device's conductance changed—essentially, how strongly it responded to signals. The device showed all the hallmarks of a real synapse: it could hold multiple levels of conductance, exhibit long-term potentiation and depression (the biological basis of learning and forgetting), and transition between short-term and long-term memory modes. When the researchers put it through learning cycles, the nano-crystalline version retained information 34.7 percent more efficiently than conventional bulk rhenium selenide.
In practical tests, the device performed edge detection on photographs and classified images from the CIFAR-10 dataset with 96.24 percent accuracy. These are standard benchmarks in machine learning, and the results suggest the technology could work in real-world applications. What makes this significant is not just the performance numbers, but the path to get there. The researchers didn't need to deposit extra layers or use complex patterning techniques. A single plasma treatment created the entire functional architecture. This simplicity matters for manufacturing at scale. Kim described the work as demonstrating a method to deliberately design the structure of van der Waals crystals for optoelectronic synapses that learn and store information using light. By solving the randomness and interface problems that plagued earlier devices, the approach opens a materials platform for neuromorphic semiconductors—chips designed to process information the way brains do. The next phase will be scaling this up and integrating it into actual AI hardware systems.
Citações Notáveis
By structurally resolving the random nature of ionic migration and interfacial issues inherent in conventional devices, this architecture can be applied to research on next-generation neuromorphic semiconductors and AI hardware.— Professor Taesung Kim, Sungkyunkwan University
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that this device mimics the structure of a biological synapse, not just its function?
Because structure determines how reliably you can control what happens. In a biological synapse, the ion channel's physical shape is what gates the flow of charged particles. If you just copy the function without the structure, you're fighting against randomness. The grain boundaries in this nano-crystalline layer do what the ion channel does—they confine and direct ionic movement at the atomic scale. That's why the device can be deterministic instead of probabilistic.
The plasma process seems almost deceptively simple. Why hasn't this been done before?
Because the obvious approach—depositing new material on top—creates interfaces that warp and degrade. This team realized they could transform what's already there. The plasma doesn't destroy the underlying crystal; it restructures only the surface. That's the insight. It's like discovering you don't need to build a new wall; you just need to rearrange the bricks in the one you have.
What does 34.7 percent better retention efficiency actually mean in practice?
It means the device forgets less. In a learning cycle, you train it, then test if it remembers, then train it again. The nano-crystalline version holds onto what it learned better than the bulk version. That's closer to how biological memory works—repeated exposure strengthens the trace. For AI hardware, that translates to fewer training cycles needed to achieve the same performance.
The 96.24 percent accuracy on CIFAR-10—is that competitive?
It's respectable, not record-breaking. But that's not the point. The point is that a device built from a single-step process, mimicking biological structure, can achieve that accuracy while processing light directly. Conventional chips separate sensing from processing. This does both simultaneously. The efficiency gain comes later, at scale.
What's the next hurdle?
Integration and manufacturing. This works in the lab on small samples. Scaling it to larger areas, integrating it with other components, and proving it uses less power than conventional approaches—that's the real test. The materials platform is proven. Now comes the engineering.