Computers have always kept thinking and remembering in separate rooms. The processor works over here; the memory sits over there. Every time one needs to talk to the other, data travels back and forth across that gap, burning time and energy at a scale that has become one of the central bottlenecks of modern AI.
A research team at DGIST in South Korea has built something that collapses that gap, and they did it using one of the smallest, most mobile elements in the periodic table.
The device, developed by Senior Researcher Lee Hyun Jun and Associate Researcher Noh Hee Yeon from DGIST’s Division of Nanotechnology, is an artificial synapse that uses electrically controlled hydrogen movement to simultaneously perform computation and store results.
The team describes it as the world’s first two-terminal AI semiconductor to achieve this using hydrogen as its active switching mechanism. The findings were published in ACS Nano.
The human brain runs on roughly 20 watts of power. It manages that by doing something conventional computers fundamentally don’t: it processes and stores information in the same place, through the same structures, at the same time. A synapse strengthens or weakens based on experience, and that change is itself the memory.
Neuromorphic computing, a field dedicated to building hardware that works more like biological neural tissue, tries to replicate that architecture. At the center of that effort is the artificial synapse, a device whose electrical conductivity changes in response to signals and holds that change after the signal stops. The more faithfully it mimics a real synapse, the more efficiently a chip can learn.
Most existing approaches to building these devices rely on oxygen vacancies, essentially defects in the material that shift under an electric field to change resistance. The method works, but it comes with persistent problems. Oxygen defects are difficult to control precisely, and their behaviour tends to drift over time, making long-term stability a chronic challenge.
“This research holds significant meaning beyond developing another AI semiconductor,” Lee said. “It presents a novel resistive switching mechanism using hydrogen migration, which is entirely different from existing oxygen vacancy-based memory.”
The DGIST device uses a stacked structure of thin semiconductor layers, with the active layer made from amorphous indium-gallium-zinc oxide, a material already common in display technology. Above it sits a silicon oxide layer that acts as both a buffer and a reservoir, and above that, a silicon nitride layer that serves as the hydrogen source.
Silicon nitride deposited through standard chip-manufacturing processes naturally contains large quantities of mobile hydrogen, a byproduct of the gases used during fabrication. That hydrogen becomes the working fluid of the device.
When a positive voltage is applied to the top electrode, hydrogen ions are pushed through the silicon oxide layer and into the active layer below. Their arrival increases electrical conductivity by raising the density of free electrons. When the voltage reverses, the hydrogen retreats. The silicon oxide layer controls how far it travels and prevents it from wandering uncontrolled when no voltage is present, which is what gives the device its memory.
Critically, this switching happens gradually, not in sudden jumps. That analog quality, where conductance rises and falls smoothly across many intermediate states, is what makes the device behave like a biological synapse rather than a simple on-off switch. It can represent not just zero and one, but a continuous range of values corresponding to synaptic weights, the numbers a neural network adjusts during learning.
“This is the first case of precisely controlling the migration of hydrogen atoms between stacked semiconductor layers electrically,” Noh said. “The findings from this study will fundamentally change the architecture of AI hardware and accelerate the era of next-generation, low-power, high-efficiency neuromorphic semiconductors.”
Reliability tends to be where promising neuromorphic devices stumble. The DGIST device held up across more than 10,000 repeated switching cycles without meaningful degradation, and retained its memory state stably for over one million seconds of continuous testing, with no measurable decline in the separation between its high and low resistance states.
The two-terminal vertical structure of the device carries its own significance. A chip architecture using two terminals rather than three reduces cell size, simplifies manufacturing, and allows far higher packing density. That matters enormously for commercial AI hardware, where the number of synaptic connections per chip determines how complex a network can be. The structure is also fully compatible with existing CMOS semiconductor fabrication processes, meaning it could in principle be manufactured using infrastructure already in place across the industry.
To demonstrate that the device could actually support learning, the researchers connected it to a simulated multilayer neural network and trained it to recognize handwritten digits using a standard benchmark dataset. Using only five discrete weight levels, the system achieved recognition accuracy above 90 percent when input images were kept at a resolution of 8 by 8 pixels or higher, a result comparable to more complex systems using many more weight states.
The finding carried a practical note: for relatively simple tasks, the number of weight levels a neuromorphic device can represent matters less than the resolution of the input data itself. More states don’t automatically mean better performance.
What separates this approach from prior neuromorphic memory research isn’t just the choice of hydrogen over oxygen. It’s the physical mechanism itself. In oxygen-vacancy devices, resistance switching happens at the interface between materials, driven by changes in barrier height. The resulting conductance changes are inherently nonlinear, which makes them harder to use as reliable synaptic weights.
In the hydrogen-based device, switching happens throughout the bulk of the active layer as hydrogen diffuses in and out. The result is a more uniform, more linear conductance change with each applied pulse, which translates directly into more predictable and controllable weight updates during neural network training.
The device also avoids one of the more stubborn failure modes in oxide-based memory: because hydrogen is supplied externally through the silicon nitride layer rather than relying on pre-existing defects in the material, the source of switching behavior is built in and renewable rather than dependent on the controlled deterioration of the active layer itself.
The power consumption demands of modern AI are significant and growing. Training and running large neural networks at scale requires data centres consuming electricity at rates comparable to mid-sized cities. Neuromorphic hardware that processes and stores information simultaneously, rather than shuttling data between separate components, represents one credible path toward AI systems that are dramatically more energy efficient.
The DGIST device’s compatibility with standard semiconductor manufacturing processes lowers one of the highest barriers between laboratory demonstration and commercial production. The materials used, silicon nitride, silicon oxide, and indium-gallium-zinc oxide, are not exotic. The fabrication steps required are already well understood across the industry.
Whether this particular architecture reaches production depends on further development and testing across a wider range of tasks and operating conditions. What the research establishes, though, is that hydrogen migration in a vertical two-terminal structure is a stable, controllable, and manufacturable basis for artificial synaptic behaviour, an option that didn’t exist before.
Research findings are available online in the journal ACS Applied Materials & Interfaces.
The original story “Introducing the world’s first AI semiconductor that thinks with hydrogen” is published in The Brighter Side of News.
Like these kind of feel good stories? Get The Brighter Side of News’ newsletter.
The post Introducing the world’s first AI semiconductor that thinks with hydrogen appeared first on The Brighter Side of News.
Leave a comment
You must be logged in to post a comment.