Researchers develop brain-mimicking AI semiconductor controlled by light

South Korean researchers have developed a brain-mimicking semiconductor that uses light to strengthen or weaken memory storage — retaining useful information while discarding the unnecessary, much like the human brain.

The Korea Research Foundation announced Wednesday that a joint research team led by professors Jo Sae-byeok and Yang Woo-seok of Sungkyunkwan University and professor Jo Jeong-ho of Yonsei University had developed a new optical synapse platform capable of independently controlling memory reinforcement and memory weakening based on the wavelength — or color — of light.

As AI systems grow more capable, their power consumption rises sharply, driving interest in neuromorphic computing, which handles memory and computation simultaneously.

Neuromorphic chips process information by mimicking the way neurons generate signals and transmit them to other neurons through synapses. Unlike conventional semiconductors, which shuttle data between separate processing and storage units, neuromorphic devices handle both functions at once — enabling fast processing of vast amounts of information at low power.

Achieving this requires artificial synapses that transmit and store information the way the brain does. Optical synapses in particular are considered a key component for next-generation AI accelerators, as they enable ultra-low-power, high-speed operation and large-scale parallel processing.

Existing optical synapses, however, have used a single switch to control both learning — the "memorizing" function — and forgetting, making it difficult to operate the two independently. Repeated learning cycles caused memory to skew in one direction, leading to overload or data loss.

The research team took a counterintuitive approach: rather than treating defects in the optical synapse material as a flaw, they turned them into a mechanism for maintaining memory stability.

The team precisely engineered microscopic defects in the metal ions within the semiconductor to create a kind of reservoir that traps electrical signals.

On top of that, they deposited a molecular layer — at nanometer-scale thickness — that selectively absorbs only near-infrared light, producing a heterojunction structure that bonds two materials with different properties.

The difference in properties created by this heterojunction structure means that the speed and direction of electron movement vary depending on the color of light applied — completing an optical synapse system in which light color alone determines whether the device memorizes or forgets.

In other words, the color of light shone on the device acts as a switch, selectively strengthening or weakening memory in the AI semiconductor.

Experiments showed that near-infrared light accelerated memorization, boosting synaptic strength by more than 13 times, while blue light released trapped electrons and rapidly weakened the signal, accelerating forgetting.

"Because learning and forgetting can both be handled using only light, the technology could be applied to low-power AI accelerators, artificial vision systems and artificial retinas," Jo Sae-byeok said. He added that commercialization would require fabricating large-scale arrays with densely integrated devices and ensuring durability under repeated operation as well as uniformity across individual devices.

The research, supported by the Ministry of Science and ICT and the Korea Research Foundation through their Outstanding Young Researcher project, was published in the international journal Nature Communications on May 18.