GWANGJU — Researchers at the Gwangju Institute of Science and Technology (GIST) have developed a multifunctional semiconductor device that can perform logic operations and artificial intelligence learning within a single element.
A research team led by Professor Kang Dong-ho of GIST's Department of Semiconductor Engineering said on the 14th that it had succeeded in selectively using two different memory functions within a single transistor to carry out both logic operations and AI learning.
The key to the research is voltage control. At high voltages, a floating-gate memory operates and sharply alters the state of current flow. At low voltages, a ferroelectric memory takes over and finely adjusts current strength. Through this approach, the team implemented different modes of computation in one device.
Using the floating-gate method, the researchers produced four operating states: insulating, metallic, n-type and p-type. Drawing on these states, they confirmed basic logic operations that either invert an input signal or pass it through unchanged. Simulating a six-transistor circuit based on actual measurements, they also showed that six logic operations were possible without rewiring the circuit.
The team implemented AI learning functions as well. Using the characteristics of ferroelectric memory, the researchers reproduced behavior in which current strength gradually strengthens or weakens, much like a synapse in the brain. When short voltage pulses were applied repeatedly, the current state of the device changed incrementally and retained that state. The gap between the lowest and highest current states was roughly 100-fold.
The team applied this property to adjusting connection strengths in a neural network. In a handwritten-digit recognition simulation, the device distinguished the numbers 0 through 9 with about 90% accuracy.
The achievement is significant in showing that a single semiconductor can serve as both a computing circuit and an AI learning element. Existing reconfigurable semiconductors have relied mainly on a single type of memory. Logic operations require broad state transitions while AI learning requires precise incremental control, which had limited efforts to implement both functions together.
The GIST team overcame that limitation by stacking atomically thin semiconductor, insulator and ferroelectric materials layer by layer. The structure divides the strengths of the two memories within a single device.
"This study demonstrates the possibility of implementing logic operations and AI learning functions together in a single device," Kang said. "It could contribute to raising the integration density and energy efficiency of next-generation AI semiconductors."
The research was supervised by Kang, with Son Hyun-chang, a combined master's and doctoral student, and Lee Seung-bin, who holds a master's degree, both of GIST's School of Electrical Engineering and Computer Science, serving as co-first authors. The findings were published online in the international journal Small on Sept. 30 of this year.