The research team led by Professor Bae Hak-yeol in the Division of Electronic Engineering, College of Engineering, Jeonbuk National University (JBNU), has succeeded in implementing a self-recovering 'brain-inspired AI semiconductor' and offered a clue to resolving a core challenge for commercialization of next-generation neuromorphic computing.

The related study was jointly conducted by the research team led by Professor Bae Hak-yeol at JBNU and the research team led by Professor Kim Tae-wan at the University of Seoul. It was published in the latest issue of the world-class journal Advanced Science (IF 14.1, JCR top 7%). (Paper title: Sustainable Synaptic Device with Two-Dimensional Ferroelectric Materials for Neuromorphic Computing).

This study is notable for elucidating the operating principle of a 'performance recovery technique' that electrically restores degraded performance occurring in synaptic transistors based on two-dimensional ferroelectric semiconductors, and for achieving improved energy efficiency at the system level.

Neuromorphic computing is a next-generation technology designed to perform computation and memory within a single device—like synapses in the human brain—and is regarded as an alternative to overcome the limitations of conventional computing architectures, such as the von Neumann bottleneck caused by separation of memory and processor.

However, existing neuromorphic devices based on memory semiconductors accumulate defects as training iterations increase, which degrades polarization switching and limits durability and long-term reliability.

To address this problem, the research team applied a 'current annealing' method using electrical pulses. They focused on the ability to control current and heat by utilizing the out-of-plane (Out-of-plane) and in-plane (In-plane) polarization characteristics that give rise to the nonvolatile behavior of ferroelectric semiconductors.

Previously, separate annealing equipment was essential to recover degradation such as defect generation in semiconductor devices caused by repeated operation and training. The current annealing method developed in this study, however, can generate controlled current and heat within the semiconductor circuit and system in a short time without external equipment, thereby restoring performance.

In fact, the research team confirmed that image classification accuracy degraded by CNN training in AI semiconductor applications was fully restored to initial levels after current annealing. In system-level simulations, an architecture was proposed that selectively recovers or deactivates only defective cells, and applicability at the large-scale array level was demonstrated.

Because this can be applied to mass-production processes for actual semiconductor chips in the future, the study secured a technical foundation applicable to chip manufacturing. Follow-up research is expected to attract significant attention in the AI semiconductor market.

Professor Bae Hak-yeol said, "The sustainability of artificial synaptic devices that perform memory and computation simultaneously is a core challenge for next-generation AI hardware. Through a self-healing technology that can be applied to existing systems without external processes, we have proposed a design direction that can maintain stable performance in long-term training and inference environments."

Meanwhile, this research was carried out with support from the Ministry of Science and ICT's Materials Global Young Connect Program, the vdW Materials and Process Technology Development Project for Ultra-High-Density Semiconductors, the Outstanding Early-Career Research Program, and the BK21 human resource training program.