Light driven neuron chip unifies sensing computing and memory

by Sophie Jenkins
London, UK (SPX) Jan 07, 2026
Modern computers consume substantial energy for tasks such as handwritten digit recognition because memory and processing are physically separated, unlike biological neurons that combine sensing, computation, and storage in one structure.

Researchers reporting in the International Journal of Extreme Manufacturing describe a neuromorphic chip that processes and learns information by tightly integrating light and electronic functions on a single platform, mimicking how neurons merge communication and memory.

The bio inspired neuron system operates at low voltages of about plus or minus 1 volt, exhibits on off current ratios up to 10^6, shows a subthreshold swing around 78 mV per decade, and maintains stable performance over 1,000 switching cycles while reaching 92.02 percent image recognition accuracy in simulations.

Corresponding author Prof. Jianwen Zhao at the Suzhou Institute of Nano Tech and Nano Bionics, Chinese Academy of Sciences, stated, "Our goal was to move away from fragmented architectures and toward a system where signal generation, transmission, and learning all happen together."

"This is much closer to how real neural systems operate."

The core hardware is a monolithic architecture built from single walled carbon nanotube thin film transistors coupled to miniature light emitting diodes, forming an electrical optical electrical loop in which electrical signals drive light emission, optical signals tune electronic behavior, and the resulting electrical outputs propagate information in a neuron like fashion.

In this design a single carbon nanotube transistor both drives the mini LEDs at about 1 volt and functions as an artificial synapse that responds to the emitted light to emulate learning behaviors.

"Combining these functions in a single device allows us to greatly simplify the circuit while improving efficiency," Prof. Zhao noted.

Unlike optoelectronic neuromorphic platforms that depend on external light sources, this chip generates and manipulates optical signals entirely on chip, with wafer scale fabrication using semiconductor compatible processes, micrometer scale features, and operating voltages around 1 volt.

When combined with an organic semiconductor layer, the carbon nanotube transistors respond across most of the visible spectrum, allowing them to process optically encoded signals created directly on the chip.

Using this integrated platform, the team demonstrated synaptic behaviors analogous to biological learning, including short term and long term modulation of signal strength.

Based on measured device characteristics, the researchers simulated a five layer convolutional neural network that exceeded 92 percent accuracy on standard image recognition tasks and achieved about 86 percent accuracy on handwritten digits without relying on separate memory or computation hardware.

The work indicates a path toward hardware in which sensing, computation, and memory are closely unified, and with further integration and scaling the approach could support compact low power neuromorphic computing, adaptive electronics, and human machine interfaces that learn and respond with reduced energy use.

Research Report: Carbon nanotube-based bio-inspired neuron systems via cascaded thin-film transistor-driven light emitting diodes and optoelectronic synaptic transistors for neuromorphic computing