Three-in-one diode integrates sensing, memory and processing for smart cameras

Think about how easily you recognize a friend in a dimly lit room. Your eyes capture light, while your brain filters out background noise, retrieves stored visual information, and processes the image to make a match. It all happens in a fraction of a second and uses remarkably little energy. Unfortunately, artificial vision systems in smartphones, cameras, and autonomous machines operate more like an assembly line. In our recent paper published in Nature Electronics, we describe how we addressed this challenge by enabling sensing, memory, and processing within the same device, pointing to a possible route toward more efficient machine vision.

The iGaN Laboratory led by Professor Haiding Sun at the School of Microelectronics, University of Science and Technology of China (USTC), in collaboration with multiple institutions, developed the multifunctional semiconductor diode with integrated photosensing, memory, and processing capabilities.

To understand the challenge, it helps to look at the basic building block of modern digital cameras: the semiconductor p-n diode. These tiny junctions act as the light-sensing pixels in imaging systems. However, a conventional diode is usually limited to a single function. It converts light into an electrical signal, and the captured data must then be transferred to separate memory and processing units. Moving this data back and forth consumes time, power, and chip area.

If engineers want to add memory or computing directly to a sensor, they often need to integrate extra electronic components around each pixel. This increases hardware complexity, takes up valuable space, and raises manufacturing cost. We wanted to find a way to overcome this single-function limit without adding extra terminals or making the device structure overly complicated.

Our solution relies on band-structure engineering. Instead of wiring together different components, we redesigned the internal structure of the diode itself. Using molecular beam epitaxy, we grew vertical semiconductor nanowires on a silicon wafer. The resulting structure contains a p-type gallium nitride layer, a middle aluminum gallium nitride layer, and a bottom n-type gallium nitride layer (Fig. 1a).

The key mechanism lies in the middle aluminum gallium nitride layer. Because this material has a wider bandgap than the surrounding layers, it forms an electron reservoir that can trap and release electrons generated by incoming light. By changing the applied voltage, we can control this charge storage behavior. This enables the same device to operate in three different modes (Fig. 1b).

At zero bias, the diode acts as a self-powered photodetector (Fig. 1c). Under a small constant bias, it behaves like an artificial photosynapse and gradually releases trapped electrons, which helps suppress random image noise in real time (Fig. 1d). With controlled voltage pulses, the stored charge can also be read out or erased, allowing the device to function as a multistate photomemory element (Fig. 1e). We found that the device could produce eight distinct electrical states. By switching among these three operating modes through voltage control, the same diode can support real-time sensing, edge noise suppression, and classification within a compact platform (Fig. 1f).

To explore its practical potential, we built a crossbar array of these diodes and applied it to a machine-learning imaging task (Fig. 2). We used the system to recognize clothing images corrupted by random background noise. In a conventional approach, a noisy image would be captured first and then sent to separate hardware for denoising and classification. In our case, the diode array performed these functions in the same hardware platform. It sensed the image, suppressed the background noise, and then used its memory states for classification.

As summarized in our study, "We show that an array of such three-in-one diodes can be used to create a compact and energy-efficient image sensor with denoising and image-classification functions without additional circuits." Using this approach, the array achieved an image recognition accuracy of more than 95%.

As artificial intelligence moves further toward edge computing for smart wearables, robotics, and autonomous systems, reducing data transfer between separate sensing and computing units becomes increasingly important. By showing that sensing, memory, and processing can occur within a single two-terminal diode, this work suggests a possible path toward more compact and energy-efficient vision hardware. We hope this band-structure engineering strategy can be developed further for future highly integrated optoelectronic systems.

This story is part of Science X Dialog, where researchers can report findings from their published research articles. Visit this page for information about Science X Dialog and how to participate.