Modern artificial intelligence has grown astonishingly capable, yet the hardware beneath it still carries the weight of an older era. Today’s computers shuttle information back and forth between memory and processors in a way that resembles a busy city with only one bridge. No matter how fast the processors become, the bridge remains a bottleneck. A research team at Daegu Gyeongbuk Institute of Science and Technology (DGIST), led by Lee Hyun Jun and Noh Hee Yeon, has taken a step toward removing that bottleneck by building a semiconductor that learns and remembers using hydrogen ions. Their work, described in the paper Tunable Hydrogen Dynamics Under Electrical Bias for Neuromorphic Memory Applications and summarized in a recent TechXplore article, offers a glimpse of a future where machines process information more like the human brain. Hydrogen is a chemical element; it has the symbol H and atomic number 1.[1]
The human brain does not separate memory and computation. Every neuron stores information in the strength of its connections, and every connection can change its strength in response to experience. This is why the brain can learn continuously and operate with remarkable energy efficiency. It performs the equivalent of massive parallel computation while consuming roughly the power of a dim light bulb. Neuromorphic semiconductors aim to capture this principle by building devices that both compute and store information in the same physical location. Instead of shuttling data across a digital highway, the device itself becomes the memory and the processor at once.
At the heart of neuromorphic hardware is the artificial synapse. This is a tiny device whose electrical conductivity can be increased or decreased in response to electrical signals, much like a biological synapse strengthens or weakens with experience. Once the conductivity changes, it stays that way until another signal alters it again. This ability to hold a state is what allows the device to store “weights,” the numerical values that represent learned information in neural networks.
For years, most artificial synapses have relied on the movement of oxygen vacancies inside oxide materials. An oxygen vacancy is simply a missing oxygen atom, and its movement changes the electrical resistance of the device. This approach works, but it comes with challenges. Oxygen vacancies are difficult to control precisely, and their movement can vary from device to device. Over time, this leads to drift, instability, and inconsistent behavior across large arrays of devices. For neuromorphic computing to scale, the underlying memory mechanism must be far more predictable.
This is where hydrogen enters the story. Hydrogen is the smallest and lightest atom, and in oxide semiconductors like IGZO it acts as a shallow donor, meaning it can easily contribute electrons and change the material’s conductivity. The DGIST team recognized that if hydrogen ions could be injected and removed in a controlled way, they could serve as a stable and reversible mechanism for memory. The challenge was that hydrogen is notoriously mobile. It tends to wander through materials unpredictably, which is why no one had previously demonstrated precise hydrogen control in a vertical, two terminal device.
The researchers solved this by designing a layered structure that acts almost like a carefully engineered garden for hydrogen. At the top sits a silicon nitride (SiNx) layer that naturally contains abundant hydrogen from its deposition process. Beneath it is a thin silicon oxide (SiOx) layer that acts as both a buffer and a reservoir. At the bottom lies the active IGZO semiconductor. When a voltage is applied, hydrogen ions drift from the SiNx layer, pass through the SiOx buffer, and enter the IGZO layer. Once inside, they form hydroxyl bonds that increase the electron concentration and lower the resistance. Reversing the voltage pushes hydrogen back toward the interface, raising the resistance again.
Think of the device as a tiny sandwich made of different layers. The researchers wanted to see where the hydrogen atoms were inside that sandwich while the device was turned on and off. To do that, they used a tool called secondary ion mass spectrometry (SIMS), which is basically an atomic level depth scanner.
Thinking of hydrogen as tiny marbles that roll around depending on how you tilt a tray makes it easier to visualize what the researchers found.
- When the device is in the SET state, which is the “on” or low resistance state, the hydrogen marbles roll down and pile up at the boundary between the SiOx and IGZO layers.
- When the device is in the RESET state, which is the “off” or high resistance state, the marbles roll back away from that boundary.
Another tool, XPS, helped the researchers see how the material’s surface chemistry changed when hydrogen was present. It showed that hydrogen fills in tiny defects in the material, almost like grout filling gaps in tile. When those gaps are filled, the material behaves more predictably and does not drift over time.
Finally, the SiOx layer acts like a sponge that slows down the movement of hydrogen when the device is not being used. Because the hydrogen does not wander off, the device can remember its state for a long time without power.
What makes this achievement especially important is that it was implemented in a two terminal vertical structure. Two terminal devices are the simplest and most compact building blocks for high density memory arrays. They can be stacked in three dimensions, allowing millions or billions of artificial synapses to be integrated into a single chip. Until now, no one had demonstrated precise hydrogen migration control in such a structure. The DGIST team’s device operated stably for more than 10,000 cycles and maintained its memory state even after long storage, a level of reliability that oxygen vacancy based devices often struggle to achieve.
The device also demonstrated analog behavior, meaning its conductivity changed gradually rather than jumping abruptly between states. This is essential for learning because neural networks adjust their weights in small increments. The team showed that the device could perform potentiation and depression, the two fundamental operations of synaptic learning, with smooth and repeatable conductance changes. In simulations, the device achieved 97.2% accuracy on the MNIST handwritten digit dataset, a strong indication that hydrogen based synapses can support real neural network workloads.
The implications of this work extend beyond a single device. By establishing hydrogen migration as a controllable and reversible mechanism for memory, the researchers have opened the door to a new class of resistive switching technologies. These technologies could enable neuromorphic chips that operate with lower power consumption than today’s GPUs and that learn continuously rather than relying on massive offline training. They could also lead to new architectures where memory and computation are intertwined, reducing the energy wasted on data movement.
In practical terms, this could influence everything from edge AI devices that need to operate on tiny power budgets to large scale AI systems that currently consume enormous amounts of energy. Hydrogen based synapses could make AI hardware more sustainable, more efficient, and more capable of real time learning.
The researchers recommend further exploration of hydrogen dynamics in oxide semiconductors and suggest that their findings could guide the design of future neuromorphic architectures. The next steps likely involve scaling the device into larger arrays, integrating it with CMOS circuitry, and exploring how hydrogen based synapses behave under more complex learning tasks.
If this line of research continues to advance, we may see a shift from today’s energy hungry AI accelerators toward hardware that learns with the elegance and efficiency of the human brain. Hydrogen, the simplest element in the universe, may turn out to be a key ingredient in the next generation of intelligent machines.
This article is shared at no charge for educational and informational purposes only.
Red Sky Alliance is a Cyber Threat Analysis and Intelligence Service organization. We provide indicators of compromise information (CTI) via a notification service (RedXray) or an analysis service (CTAC). For questions, comments or assistance, please contact the office directly at 1-844-492-7225, or feedback@redskyalliance.com
Weekly Cyber Intelligence Briefings:
- Reporting: https://www.redskyalliance.org/
- Website: https://www.redskyalliance.com/
- LinkedIn: https://www.linkedin.com/company/64265941
Weekly Cyber Intelligence Briefings:
REDSHORTS - Weekly Cyber Intelligence Briefings
https://register.gotowebinar.com/register/5207428251321676122
[1] https://six3ro.substack.com/p/hydrogen-as-a-new-pathway-for-learning
Comments