| Jun 20, 2026 |
Light color controls a photonic synapse that remembers and forgets
Researchers built a photonic synapse that strengthens or erases memory by light color, using a defect to mimic the brain's balanced learning.
(Nanowerk News) Researchers at Sungkyunkwan University in South Korea have built a photonic synapse, an artificial synapse controlled by light, whose memory can be strengthened or erased by changing the color shining on it. The device locks in information under near-infrared light and rapidly weakens that memory under blue light, reproducing the brain's ability to keep learning in balance.
|
|
The work, published in Nature Communications in Nature Communications ("Disorder-mediated non-equilibrium photocurrent redistribution enables homeostatic synaptic conditioning in AgBiS2 heterostructure"), depends on a material defect that engineers usually try to remove.
|
Key Findings
- Near-infrared light strengthened the synaptic connection more than 13-fold, while blue light rapidly weakened it.
- The effect comes from controlled disorder in silver bismuth sulfide, an imperfection that traps electrons and retains memory after the power is removed.
- In a digit-recognition test, the device kept learning stably over 1,000 training rounds, compared with about 200 rounds for a conventional single-mechanism approach.
|
|
Modern artificial intelligence consumes enormous amounts of power. Training a single generative model can draw as much electricity as a small city. The brain does far more on less energy than a light bulb because it stores and processes information in the same place, the synapse.
|
|
That efficiency has drawn researchers toward neuromorphic, or brain-inspired, computing, and especially toward photonic synapses that use light for fast, ultralow-power operation.
|
|
Conventional artificial synapses carry a built-in weakness. They use one control to handle both remembering, known as potentiation, and forgetting, known as depression. Over time the balance breaks down. The stored weights either run away to saturation or fade to nothing, erasing what the device had learned. Biological brains avoid this through homeostatic plasticity, but artificial hardware has had to copy that behavior with extra software that adds cost.
|
|
The team led by Professor Sae Byeok Jo and Professor Wooseok Yang took the opposite approach and used the imperfection rather than removing it. In silver bismuth sulfide (AgBiS2), a light-absorbing semiconductor being explored for next-generation devices, a small and controlled disorder in the arrangement of ions creates traps that hold light-generated electrons for long periods.
|
|
That trait is a drawback for fast detectors. But it lets the material act as a natural memory that keeps information even after the power is switched off.
|
|
By tuning this disorder and adding a near-infrared-absorbing molecular layer on top, the researchers turned the color of incoming light into a learning switch. Near-infrared light produced accelerated learning, boosting the synaptic connection more than 13-fold. Blue light drove accelerated forgetting and quickly weakened it.
|
|
Using ultrafast laser spectroscopy able to resolve events down to a quadrillionth of a second, the team confirmed that the two colors send electrons along opposite routes, filling the traps or emptying them.
|
|
The advantage showed up in a handwritten-digit recognition simulation. Neural networks built on a single mechanism lost their memory within about 200 training rounds. The new scheme, which assigns separate wavelengths to learning and forgetting, kept recognizing patterns reliably across more than 1,000 rounds, brain-like learning that held up at the hardware level.
|
|
Professor Jo explained the reasoning behind the design. "Knowing how to forget is as important as knowing how to remember. The essence of this work is that we separated those two functions by the color of light, and revived what was considered a defect into a self-balancing learning function for AI hardware," he said.
|
|
The method is not tied to a single material, and the entire process uses low-temperature, ink-based solution techniques that fit existing semiconductor production lines. The researchers expect it to support light-based neuromorphic computing, low-power AI accelerators, in-sensor computing, and machine-vision systems for autonomous vehicles and robots, along with artificial retinas that can both see and remember.
|
