Atomically thin transistor stores 3,024 states
By Engineers at Nanjing University of Aeronautics and Astronautics have unveiled an atomically thin ferroelectric transistor capable of storing 3,024 distinct polarization states – a breakthrough that could power energy-efficient neuromorphic electronics and edge AI accelerators. The device is built from layered materials including graphene and hexagonal boron nitride (hBN), leveraging a moiré pattern to precisely control its states through electrical pulses.
The transistor demonstrated stable states for over 100,000 seconds and achieved 93% accuracy in image recognition tests. The researchers emphasize the device’s simple structure and scalability potential, though performance improvements and industrial process integration are still needed.
How an atomically thin ferroelectric transistor differs from conventional ones
The key difference lies not in the usual binary on-off operation, but in its analog, multilevel polarization. Sliding layers of graphene and hBN alter the local electric field, producing thousands of stable levels, whereas most existing analog memory elements offer only a few dozen states or operate in digital mode.
- 3,024 stable polarization states;
- Stability for over 100,000 seconds;
- 93% accuracy in image recognition tests by the team;
- Constructed from graphene + hexagonal boron nitride (hBN) and moiré structures;
- Operates at room and elevated temperatures; simple structure promises scalability.
When can we expect this transistor in chips?
Lab success is just the beginning. For commercial adoption in neuromorphic modules, the technology must pass three major hurdles: switching speed and cycling endurance, integration with CMOS processes, and defect-free mass production of heterostructures. Many of these challenges are familiar to the industry from the development of FeFETs based on ferroelectric HfO2 and memristors; transitioning from prototypes to fab lines often takes years and requires solving materials compatibility and quality control issues.
Still, the core idea-a multilevel memory cell in a single element-addresses practical issues like energy consumption and interface overhead between analog “synapses” and digital neural networks. If response times can be improved and endurance ensured, these transistors could first appear in specialized edge AI modules, eventually expanding to broader neuromorphic platforms.
The open question remains whether the moiré-controlled advantages will hold up in mass production and how quickly the industry can adapt layered crystal fabrication to cost targets. The outcome will determine if this discovery becomes a viable alternative to existing solutions or remains a laboratory showcase of twistronics’ potential.
