Solved the Greatest Challenge of AI Semiconductors... Ultra-Low-Power Neuromorphic Achieved Without Selector Devices [Reading Science]

Simultaneous Solution to Nonlinearity and Leakage Current in a Single Device
GIST, Seoul National University, and USC Collaborative Research
Laying the Foundation for Accelerated Commercialization of Highly Integrated AI Semiconductors

A new neuromorphic semiconductor device that can increase both the integration density and process efficiency of next-generation artificial intelligence (AI) semiconductors has been developed. The domestic research team has successfully implemented both memory functionality and self-rectifying—the ability to allow current to flow in only one direction—within a single device. As a result, stable information storage and processing are now possible even without the separate selector component that was previously required.

Structure and rectification characteristics of the ferroelectric-based single device neuromorphic crossbar array developed by the collaborative research team. (From left) Schematic diagram of the device structure, microscope image of the fabricated device, and electrical characteristic graph showing the magnetic 'rectification' property where current flows in only one direction. Provided by the research team

Gwangju Institute of Science and Technology (GIST) announced on July 30 that the research team led by Professor Sanghan Lee of the Department of Materials Science and Engineering, together with the team led by Professor Ho-won Jang of the Department of Materials Science and Engineering at Seoul National University and the team led by Professor Joshua Yang of the Electrical and Computer Engineering Department at the University of Southern California (USC), has jointly developed a neuromorphic semiconductor device that incorporates both memory and self-rectifying properties into a single unit.

Neuromorphic semiconductors are next-generation AI chips that process information storage and computation simultaneously, just like the human brain. Utilizing a crossbar array structure that connects memory elements in a grid pattern, they can process massive amounts of data in parallel, making this architecture a key technology for implementing ultra-low-power AI systems.

However, conventional neuromorphic semiconductors have faced significant technological hurdles: nonlinearity, where information is not stored consistently under repeated electrical stimuli, and leakage current that causes current to spill over into neighboring devices. To overcome these, individual selector components had to be added to each memory cell. As a result, the device structure became more complex, leading to reduced integration density and lower process efficiency.

Integrating Memory and Selector Functions into One

The research team precisely controlled the migration of oxygen vacancies by adding barium (Ba) to bismuth ferrite (BiFeO₃), a ferroelectric material capable of retaining electrical signals. This approach enabled them to implement linear information storage characteristics—preserving information at a consistent magnitude even under repeated stimuli—while at the same time realizing self-rectifying functionality, which allows current to flow in only one direction, directly within the device.

This technology effectively blocks leakage current without the need for separate selector components, thereby simplifying the circuitry while ensuring stable operation. According to the research team, eliminating the selector components can also increase semiconductor integration density and simplify the manufacturing process.

Group photo of the GIST-Seoul National University-USC international joint research team. (From left) Howon Jang, Professor of Materials Science and Engineering at Seoul National University; Sanghan Lee, Professor of Materials Science and Engineering at GIST; Joshua Yang, Professor of Electrical and Computer Engineering at the University of Southern California (USC); Youngmin Kim, Integrated MS-PhD student in Materials Science and Engineering at Seoul National University; Yunjung Lee, Professor of Electronic Engineering at Kookmin University; Jiwoong Yang, Integrated MS-PhD student in Materials Science and Engineering at GIST. Provided by GIST

The research team also obtained meaningful results from performance verification. The newly developed device maintained its stability through more than 10 million repeated write/erase operations. They implemented an 11×11 crossbar array (containing 121 memory cells) and demonstrated stable information storage and control of text and image patterns during storing and erasing experiments, all without the aid of selector devices. Notably, the rectification ratio for suppressing leakage current improved by over one million times.

Sanghan Lee, Professor at GIST, stated, "This study is academically and industrially significant, as it solves both the nonlinearity of information storage and the leakage current issues—previously key challenges in neuromorphic semiconductors—within a single device." He added, "It presents a new design paradigm that could accelerate the commercialization of high-performance, ultra-low-power AI semiconductors as well as advance efficient neuromorphic computing."

This research was supported by the Ministry of Science and ICT and the National Research Foundation of Korea through the Individual Basic Research Program (Mid-Career Researcher Program), the Nano and Material Technology Development Program, and the Future Materials Discovery Program. The results were published online in the international journal Nature Communications in March this year, with the final corrected version released this month.

This content was produced with the assistance of AI translation services.