On November 10, the Chinese Academy of Sciences announced a nitride semiconductor device that remembers an electrical state produced by light input and can switch to fast detection when the voltage is changed. The mechanism separates photogenerated charges at the boundaries between semiconductor layers and controls how charges move through the metal contacts. A single device can therefore switch between a memory function, which holds charge for a long time, and a fast-detection function, which extracts charge quickly. The researchers also reported an image-classification accuracy above 90%, but that figure comes from a computational model that incorporates measured device data. It should be kept separate from the ultraviolet detection and memory performance confirmed on the actual device.

The work was carried out by a joint research team that includes the Changchun Institute of Optics, Fine Mechanics and Physics of the Chinese Academy of Sciences. The authors include Ke Jiang, Bingxiang Wang, and Xiaojuan Sun, with Dabing Li as corresponding author. The paper was published in Nature Materials on October 27 (DOI: 10.1038/s41563-026-02780-7).

AD

Separating photogenerated charges and controlling retention with metal contacts

The team used a structure that stacks layers of aluminum gallium nitride (AlGaN) with different ratios of aluminum to gallium. Polarization charges arise at the boundaries between layers of different composition, creating an internal electric field. This "interface polarization" is used to move the electrons and holes generated by light to different interfaces. A hole is the place where an electron is missing, which behaves like a positively charged particle.

According to the Chinese Academy of Sciences, conventional charge-retention methods that rely on material defects have a weakness: electrons and holes recombine, charge is lost, and the response weakens. In the new structure, the internal field separates electrons and holes in space, which suppresses losses from recombination. The team also used nickel (Ni) and titanium (Ti) to form asymmetric metal contacts on the left and right, and the applied voltage determines whether holes are held at the interface or released. What the device remembers is not the light itself but the electrical state changed by the light input.

To confirm the mechanism, the team ran control experiments with different semiconductor layer structures. They compared devices with 0, 1, and 3 channels under identical conditions: a forward voltage of 1.5 V, ultraviolet light at 255 nm, a light intensity of 10 μW/cm², and an irradiation time of 5 seconds. The 0-channel and 1-channel devices, which lack a negatively polarized interface, showed no notable retention. The 3-channel device kept its current flowing after the light was switched off.

The team also compared metal contact configurations. Using the same fabrication process, they made devices with asymmetric Ni/Ti contacts and devices with symmetric contacts using Ni on both sides. The symmetric contacts did not give long-term nonvolatile retention. Varying the semiconductor layers and the contacts separately supports the effectiveness of a design that combines charge separation with controlled charge movement.

The team also analyzed the internal electric field and charge distribution with APSYS simulations. The 30% polarization level used in the calculations is a setting that accounts for screening of interface charge. Separate material measurements on two samples examined charge distributions corresponding to the differences in composition. These calculations and measurements are meant to clarify the internal charge distribution, and play a different role from quantitatively predicting device performance.

Switching voltage to combine light memory and fast detection

When a forward voltage is applied, the device holds the state produced by light input. At zero or reverse voltage, it switches to fast detection mode. With the same semiconductor structure, the two functions are selected by changing how charges move through the contacts. According to the Chinese Academy of Sciences, this operation was demonstrated on the same prototype device.

Applied voltage Hole behavior Function
Forward Held at the interface Remembers the electrical state produced by light input
Zero or reverse Releases the held holes Detects light at high speed

Note: This shows the general correspondence of operation. Actual applied voltages and illumination conditions vary by measurement.

In this device, the memory behavior, in which current persists after illumination, and the fast response can be selected by the applied voltage. The main target is ultraviolet light, with a peak response wavelength of 235 nm and a long-wavelength cutoff of 275 nm. This is a different wavelength range from that of ordinary cameras that capture visible light. The paper also reports sensitivity to very faint light, on the order of nanowatts per square centimeter, in memory-sensor mode.

Fast response was evaluated with a 193 nm ArF pulsed laser. The repetition frequency was 200 Hz and the pulse width about 7 ns. Measurements with an amplifier and oscilloscope gave a rise time of 8.35 ns and a fall time of 7.81 ns. A different readout configuration shown in the supplementary material, however, has transient responses on the millisecond scale. The authors explain that this difference is mainly due to limitations of the external circuit.

Heat resistance was evaluated under conditions separate from the fast-response tests. Using an LED at 255 nm and 10 μW/cm², the team measured characteristics from 50 to 300 °C, applying 0 V in detection mode and +1.6 V in retention mode. In repeated tests at 300 °C, one cycle consisted of writing with a 0.1-second light pulse and resetting by applying −30 V for 1 ms, and this was repeated for 1,000 cycles. After each operation, 100 data points were collected, and their average was used to evaluate the state.

Performance did not, however, stay constant at high temperature. In the repeated tests at 300 °C, the switching ratio varied from about 2.1×10^4 to about 1.1×10^7. Dark current also rose with temperature, and specific detectivity, which expresses the ability to distinguish faint light from noise, declined. A response to light was still confirmed at 350 °C, but retention stability fell.

AD

Over 90% image-classification accuracy, plus a long-term retention test on 81 devices

The team also examined whether the device's characteristics could be used for image recognition. The classification accuracy above 90% was not obtained by running image recognition on an actual device array. It came from a simulation in which measured current states were built into a neural network. The study examined whether the electrical states the device can hold could be used for computation.

The evaluation used Fashion-MNIST, a dataset of clothing images. The task is to classify 28×28-pixel grayscale images into 10 categories. The 36 discrete current levels that were actually measured were normalized and mapped to the weights of the neural network. The model consists of three convolutional blocks and two fully connected layers, with weights restricted to 36 levels. The supplementary material shows results from 30 epochs of training, but the number of images used for training and for evaluation could not be confirmed.

The Chinese Academy of Sciences says that the classification accuracy of the model using 36 current levels is close to that of a software model whose weights are not restricted to such discrete values. No specific figure for the difference in accuracy between the two could be found. The accuracy above 90% is a result from a computational model that reflects measured device characteristics.

The team did fabricate actual device arrays, three in total. For one of them, with 81 devices, statistical data are presented that track the switching ratio after storage in air for one month and for one year. This was a hardware test across multiple devices that examined whether they could detect ultraviolet light and retain the state, and what it evaluates differs from the neural-network image classification.

The team's goal is application to in-sensor computing, which reduces the burden of transferring raw data between photodetectors, memory, and processors. This study demonstrated that detection and memory of light can be performed in the same device, and also evaluated a computational model that uses its electrical states for image classification. The results have been peer reviewed and control experiments varying the structure and contacts were carried out, but as of November 13, 2026, no independent replication has been confirmed. Integration with existing nitride semiconductor manufacturing technology and CMOS circuits is also part of the future applications the team envisions.

To be used as an actual computing system, the device will need to be evaluated in large-scale hardware that combines device arrays with readout circuits, looking at power consumption, integration density, manufacturing yield, and more. A key question for practical use is how much the memory function can reduce the data-transfer burden while the nanosecond-scale response is preserved across the whole circuit.