A layered transistor design combines light detection optical memory and neuromorphic processing in one unit offering compact and efficient artificial vision hardware.
(Nanowerk Spotlight) Artificial vision systems rely on three separate modules: sensors to detect light, memory units to store visual input, and processors to interpret the information. Designers typically separate these components across different physical units, which increases system complexity, power consumption, and latency. This architecture stems from the conventional von Neumann model, which splits memory and computation into distinct blocks. While this model supports flexibility, it limits the efficiency of data exchange—a key drawback in applications that require real-time visual processing.
Engineers have attempted to close this gap by developing dual-function components. Some optoelectronic devices can perform sensing and limited processing or memory functions. Others emulate basic synaptic behavior. But most of these devices rely on complex layering, suffer from unstable interfaces, or underperform in endurance, sensitivity, or speed. Efforts to fully combine sensing, storage, and computation in a single component remain rare.
A research team at South China Normal University and Guangdong University of Technology developed a neuromorphic transistor that integrates all three functions. Their device combines photo-sensing, non-volatile memory, and synaptic computation in one layered unit. It responds to light, stores visual input, and processes signals—all controlled through a gate voltage. This integration aims to simplify the design of artificial vision systems and reduce their energy and hardware overhead.
Device concept, structure, and characterization. a) Schematic diagram of the functional and application demonstration of the device operating at different modes. b) Schematic diagram of the transistor. c) Optical micrograph of the fabricated device. The scale bar is 10 μm. d) Height distribution of component materials in the fabricated device measured by AFM. e) Raman spectra of components in the fabricated device. f) Cross-sectional highresolution transmission electron microscope (TEM) image of the constituent layers in our fabricated device. g) Scanning TEM image and corresponding energy-dispersive X-ray spectroscopy (EDS) elemental mapping image of the fabricated device. (Image: Reprinted with permission by Wiley-VCH Verlag) (click on image to enlarge)
The team built the transistor using a vertical stack of two-dimensional materials. MoS₂ serves as the transport channel. MoTe₂, a light-sensitive semiconductor, sits above it and interacts directly with incoming photons. An h-BN dielectric layer and a graphene gate electrode complete the stack. By adjusting the voltage applied to the gate, the researchers controlled the electric field across the MoTe₂/MoS₂ interface. This field dictates how the device functions, allowing dynamic reconfiguration between three modes.
At a negative gate voltage, the device operates as a light detector. When exposed to light, MoTe₂ generates electron-hole pairs. The internal electric field drives electrons into the MoS₂ layer while trapping holes in MoTe₂. This trapped charge alters the electric landscape and amplifies current through MoS₂—a phenomenon known as the photo-gating effect. The researchers measured a responsivity of 6,515 amperes per watt and a detectivity of 3.92 × 10¹⁴ Jones, which indicates the ability to sense faint optical signals with minimal background noise.
This same mechanism enables optical memory. The device retains current after the light source is removed because the trapped holes in MoTe₂ persist. The researchers measured this persistent photoconductivity for over 10,000 seconds without any additional power input, confirming the device’s non-volatile behavior. A short pulse of positive gate voltage releases the stored charge and resets the device. Each write or erase operation completes in under a millisecond, and the system remained stable through more than 100 full cycles.
By adjusting the number and intensity of light pulses, the researchers programmed multiple distinct memory levels. They recorded 19 discrete storage states under low-voltage conditions, enabling multi-bit memory in a single unit. The readout current scaled linearly with the number of light pulses, making the device suitable for analog data storage and easy neural network integration.
At zero or positive gate voltage, the device shifts to a synaptic mode. It responds to repeated light pulses in a way that mimics synaptic plasticity—the ability of biological synapses to change strength based on stimulus history. The team demonstrated paired-pulse facilitation: two closely spaced pulses produced a larger response than a single one.
Repeated pulses led to long-term potentiation, with gradually increasing output that decayed over time once stimulation stopped. When the researchers applied the same stimulation again, the device regained its prior state faster, modeling the biological process of relearning.
The team used the excitatory postsynaptic current (EPSC) to characterize these behaviors. They varied light pulse timing, duration, and intensity to control the strength and duration of the synaptic response. By tuning the gate and drain voltages, they also adjusted the sensitivity and speed of synaptic transitions. These parameters let the device simulate a wide range of neural dynamics using a single configuration.
To evaluate its practical use, the researchers applied the device to an artificial neural network. They trained the network on the MNIST dataset of handwritten digits, using the transistor’s conductance states to represent synaptic weights. They increased weights with optical stimulation and decreased them with electrical pulses. The network achieved 95.26% accuracy on the training set and 94.29% on a separate test set. This performance, close to ideal software models, shows that the device can support robust learning in real neural network tasks.
The transistor operates at low voltage—just 0.1 volts for most synaptic behavior—and does not require external memory or processors. Its compact design reduces the footprint of artificial vision systems and simplifies their architecture. Because the materials are thin and flexible, the device can be integrated into bendable substrates for wearable or embedded electronics.
This work demonstrates that a single, reconfigurable transistor can replicate three core functions of visual processing—detection, memory, and interpretation. The researchers used electrostatic doping to control charge movement across layered 2D materials, combining separate computational roles in one physical structure. This approach could lead to smaller, more efficient artificial vision platforms that operate closer to the way biological systems function, using local, parallel processing instead of distant modular hardware.
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
