A new study published in Nature Materials reports that researchers have developed an aluminium gallium nitride device that combines light detection, memory and basic processing by controlling charge at polarization interfaces. The device can detect light at the nanowatt-per-square-centimeter level and operate stably at temperatures up to 300 °C.

The study was led by researchers from the Changchun Institute of Optics, Fine Mechanics and Physics of the Chinese Academy of Sciences (CAS), Peking University, East China Normal University, the Shanghai Institute of Technical Physics of CAS and Zhejiang University.

In-sensor computing reduces the need to transfer raw data between separate detectors, memory units and processors. Existing devices often rely on defects to store charge, which can consume photogenerated carriers and weaken the response. Other approaches use stacked two-dimensional materials that remain difficult to manufacture over large areas. Group III nitride semiconductors offer an alternative because their material quality, fabrication processes and industrial supply chain are well-established.

The new device uses an aluminium gallium nitride heterostructure made of alternating layers with different compositions. These changes create interfaces with opposite polarized charges and built-in electric fields. These fields drive photogenerated electrons and holes toward different interfaces, reducing their recombination. Asymmetric nickel and titanium contacts then control the movement of the separated carriers. Under forward bias, holes are retained at one interface while electrons produce a readable current, allowing the incident light to be stored. Under zero or reverse bias, the stored holes are released and both carrier types move rapidly, switching the same component into a detector.

This control scheme provides the fabricated device with multiple operating modes without changing its physical structure. It can record a persistent optical state after bright illumination, form multiple temporary states under weaker pulses and return to its initial condition through an electrical reset. In detector mode, it responded on the nanosecond timescale. The device also maintained a large current contrast and repeatable write-erase behavior after extended storage in air.

To evaluate the device's computing potential, the researchers used experimentally measured device states in a neural network simulation for clothing image classification. The device-based model achieved more than 90 percent accuracy and remained close to the corresponding full-precision software model. Tests on a fabricated device array further showed that individual pixels could detect and retain optical inputs, supporting the use of the material platform for parallel sensing and local preprocessing.

The study also clarifies the current limits. It demonstrates individual devices and a small array rather than a complete large-scale computing system. The image classification result was obtained through a network simulation parameterized by experimental measurements. Therefore, system-level energy consumption, integration density, readout circuitry, and manufacturing yield still require direct evaluation in larger hardware implementations.

The results identify interfacial polarization as a practical design tool for integrating sensing and memory in a mature semiconductor family. Since the device uses one nitride material system and standard fabrication steps, this approach could simplify future integration with established semiconductor processes.