The grain boundaries confined ionic migration, giving researchers deterministic control
In a South Korean laboratory, researchers have coaxed a semiconductor crystal into behaving like the junction between two neurons — not through complexity, but through a single chemical step that lets the material's own architecture do the work. The device, built from layered van der Waals crystals, responds to light the way a synapse responds to neurotransmitters, learning, forgetting, and relearning with measurable fidelity. At a moment when artificial intelligence demands ever more power to see and interpret the world, this work offers a quieter path: hardware that computes not by force, but by resemblance to life.
- Conventional AI chips consume enormous power processing visual data, and the field urgently needs materials that compute more like brains and less like brute-force calculators.
- Van der Waals crystals held promise for neuromorphic hardware for years, but unpredictable grain boundaries, polymer residue, and structural degradation made precise manufacturing nearly impossible.
- A single-step plasma sulfurization process transformed rhenium selenide crystals into a two-layer structure that naturally mimics biological ion channels — no extra deposition, no patterning, the architecture emerging from chemistry alone.
- Grain boundaries in the nano-crystalline surface layer channel ionic migration along precise pathways, giving researchers deterministic control over synaptic weight updates — the core mechanism of learning.
- The device achieved 96.24% accuracy on a standard image classification benchmark and retained memory 34.7% more efficiently than bulk material, signaling readiness as a foundation for real neuromorphic systems.
- The immediate horizon is scaling this materials platform and integrating it into larger hardware architectures capable of real-time visual processing at biological efficiency.
At Sungkyunkwan University in South Korea, Professor Taesung Kim's team has built a light-sensitive crystal that performs the work of a synapse — the junction where neurons communicate — using photons instead of neurotransmitters. The device is constructed from layered van der Waals materials, atomically thin and optically responsive, but long considered too difficult to manufacture reliably at scale. Grain boundaries formed unpredictably, polymer residue accumulated, and crystalline structure degraded across large areas, making precise control of the material's response to light nearly impossible.
The team's solution drew from biology. Light-sensitive ion channels in cell membranes act as selective barriers gating the flow of charged particles. Kim's team asked whether semiconductor material could be made structurally analogous. Their answer was a single-step process: exposing bulk rhenium selenide crystals to a plasma of argon and hydrogen sulfide. Sulfur atoms migrated into the upper surface, creating a nano-crystalline layer of tiny grains while leaving the bulk material beneath pristine and single-crystalline. The two-layer structure — nano-crystalline surface atop intact bulk — emerged from the chemistry itself, requiring no additional fabrication steps.
The grain boundaries between nano-sized surface grains proved to be the key. Scanning probe microscopy revealed that sulfur ions moved through the material along pathways confined by those boundaries, giving researchers deterministic control over how the device's conductance changed with light exposure. This is synaptic weight update — how a connection strengthens or weakens. The device demonstrated multi-level conductance, long-term potentiation and depression, paired-pulse facilitation, and tunable transitions between short- and long-term memory. In learning-forgetting-relearning cycles, it retained information 34.7% more efficiently than conventional bulk material.
In practical tests, the device performed edge detection on natural images and reached 96.24% accuracy on the CIFAR-10 image classification benchmark. The results position the material as a viable building block for neuromorphic hardware — systems designed to process visual information in real time with minimal power consumption. The findings appeared in Advanced Materials on June 3, 2026, with support from South Korea's National Research Foundation and collaborators across three institutions. The next challenge is scaling the platform and integrating it into larger systems where brain-inspired computing can be tested against the full demands of the real world.
In a laboratory at Sungkyunkwan University in South Korea, researchers have engineered a crystal that responds to light the way a brain cell responds to chemical signals. The device, built from layered van der Waals materials, performs the essential work of a synapse—the junction where neurons communicate—but does it with photons instead of neurotransmitters. The breakthrough comes from Professor Taesung Kim's team, who found a way to manufacture these light-sensitive crystals using a single-step process that had eluded the field for years.
The problem they solved was structural. Conventional van der Waals materials, which are atomically thin and optically responsive, had proven difficult to manufacture at scale. Grain boundaries would form unpredictably. Polymer residue would accumulate during fabrication. Interfaces would warp. The crystalline structure would degrade across large areas. These defects made it nearly impossible to control how the material responded to light, which is essential if you want to use it as a computing component. The researchers approached the challenge by looking at nature. Light-sensitive ion channels in biological cell membranes have a particular architecture—a selective barrier that gates the flow of charged particles. Kim's team asked whether they could build something structurally analogous from semiconductor material.
Their solution involved exposing bulk rhenium selenide crystals to a plasma mixture of argon and hydrogen sulfide. The process took a single step. The sulfur atoms migrated into the upper surface of the crystal, transforming it into a nano-crystalline layer composed of tiny grains, while leaving the bulk material beneath untouched and single-crystalline. The result was a two-layer structure: a nano-crystalline surface layer corresponding functionally to a biological ion channel, sitting atop a pristine bulk layer corresponding to the cell's interior. No additional deposition. No patterning. The architecture emerged from the chemistry itself.
What made this architecture work was the grain boundaries. Using scanning probe microscopy, the researchers traced how sulfur ions moved through the material. The boundaries between nano-sized grains confined this ionic migration to precise pathways, giving the researchers deterministic control over how the device's conductance changed in response to light. This is the synaptic weight update—the mechanism by which a synapse strengthens or weakens a connection. The device demonstrated the full repertoire of synaptic behaviors: it could hold multiple levels of conductance, exhibit long-term potentiation and depression (the biological basis of learning and forgetting), show paired-pulse facilitation (a form of short-term plasticity), and transition between short-term and long-term memory modes. When the researchers ran learning-forgetting-relearning cycles, the nano-crystalline device retained information 34.7 percent more efficiently than conventional bulk material.
In practical tests, the device performed edge detection on natural images and achieved 96.24 percent accuracy on the CIFAR-10 image classification benchmark—a standard test in machine learning. These results suggest the material could serve as a building block for neuromorphic hardware, systems designed to process visual information the way biological brains do, in real time and with minimal power consumption. The research was supported by South Korea's National Research Foundation, the Institute for Basic Science, and the Ministry of Trade, Industry and Energy, and involved collaboration across three institutions. The findings appeared in Advanced Materials on June 3, 2026.
What matters about this work is not that it mimics biology for its own sake. It matters because artificial intelligence systems increasingly need to process vast streams of visual data—surveillance feeds, autonomous vehicle sensors, medical imaging—and do it quickly and efficiently. Conventional computer chips handle this through brute computational force, which consumes enormous amounts of power. A neuromorphic approach, using materials that naturally compute the way neurons do, could change that equation. This crystal is a materials platform, a foundation. The next step is scaling it, integrating it into larger systems, and testing whether it can deliver on the promise of brain-inspired computing in the real world.
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 the structure matter so much? It's still just a crystal responding to light.
Because structure determines function. The grain boundaries act like gates. They control where the sulfur ions can move, which controls how the electrical conductance changes. Without that control, you have randomness. With it, you have a device that behaves predictably, like a synapse.
And the biological analogy—is that just poetic, or does it actually guide the engineering?
It guides it. The researchers looked at how ion channels work in cell membranes and asked: what's the essential architecture? Then they built it from semiconductor material. The nano-crystalline layer and the bulk layer below it structurally correspond to the channel and the cell interior. That correspondence is what makes the device work.
The 34.7 percent improvement in retention—what does that mean practically?
It means the device forgets more slowly during learning cycles. If you're training a system to recognize images, this device would need fewer repetitions to lock in what it learned. That translates to less power consumption, faster training.
And the 96.24 percent accuracy on CIFAR-10—is that competitive with conventional chips?
It's respectable, but this isn't about beating conventional chips at their own game. It's about doing the same work with a fraction of the power, in a form factor that could be integrated directly into sensors. The real competition isn't speed. It's efficiency.
What's the next hurdle?
Manufacturing at scale. They've proven the concept works. Now they need to make thousands of these devices reliably, integrate them into larger systems, and test whether the efficiency gains hold up in real applications. The lab is one thing. The factory is another.