The rise of intelligent edge applications, such as micro-robots and wearable devices, has imposed stringent requirements for low power consumption in visual perception systems. Although the traditional complementary metal-oxide-semiconductor (CMOS) architecture is technologically mature, the computational and memory overheads associated with analog-to-digital conversion and algorithm execution often exceed the extremely low power budgets of edge devices. In-sensor computing technology integrates optical sensing and analog computation within a single chip, reducing data redundancy and significantly lowering energy consumption from the source, making it a key pathway to address the above bottleneck. Currently, researchers often attempt to use electrostatic doping or two-dimensional materials to encode tunable photoresponsivity weights, thereby performing the matrix-vector multiplication operations at the core of neural networks.

Current direct photocurrent computation still faces fundamental challenges such as non‑robust device physics and difficulty in integration scaling. Many existing schemes lack true bipolar, non‑volatile photoresponsivity, and the high‑temperature environment required for integrating two‑dimensional materials can easily damage the critical readout circuit dielectric. The reliance on manual transfer techniques also leads to low manufacturing yield. Moreover, as the complexity of vision tasks increases, analog‑domain processing faces efficiency issues related to superlinear energy scaling. Striking a balance between operational simplicity and analog computation overhead, while achieving large‑scale monolithic integration with mature readout circuits, remains a key bottleneck restricting the practical application of in‑sensor computing systems.

To address the above issues, a team from Beijing Institute of Technology, Sun Yat‑Sen University, Xiamen University, Hong Kong University of Science and Technology, and other institutions conducted systematic research using ZEPTOOLS' in‑situ TEM measurement system. The team developed a symmetry‑reconfigurable photodiode based on AgBiS₂ material. By triggering localized electrochemical reactions of silver ions through voltage programming, they achieved reversible switching of device symmetry, thereby providing bipolar photoresponsivity weights with non‑volatility and high scaling potential for in‑sensor direct photocurrent computation.

Title: A symmetry-reconfigurable photodiode for sensing and computing
Journal: Nature Electronics
URL：https://doi.org/10.1038/s41928-026-01617-0

This study proposes and constructs, for the first time, a symmetry‑reconfigurable photodiode based on the AgBiS₂ material, aiming to fundamentally solve the basic physical challenge of single weight in direct photocurrent computation for edge hardware. The researchers deposited micro‑/nano‑structured arrays of the broadband light‑absorbing material AgBiS₂ between high‑work‑function metal electrodes, forming back‑to‑back symmetric Schottky junctions. In the unbiased state, this symmetric structure effectively blocks dark current. When a specific programming voltage is applied, localized electrochemical reduction of silver ions within the material occurs, lowering the Schottky barrier on one side and thereby breaking the original symmetry. This physical process endows the device with the ability to switch freely between positive and negative short‑circuit photocurrents, providing true bipolar, non‑volatile optoelectronic response weights for neuromorphic vision systems, overcoming the limitations of complex conventional device architectures.

To conclusively verify the microscopic physical origin of the above symmetry reconfiguration, the research team conducted in‑depth dynamic characterization at the nanoscale using in‑situ electron microscopy. In the transmission electron microscopy experiments observing cation migration, the researchers specifically introduced the ZEPTOOLS in‑situ sample holder to provide stable real‑time electrical signal stimulation and a precise electrical testing environment for the microscopic chip samples. Through the precise manipulation and data acquisition enabled by this in‑situ sample holder, the study not only clearly captured the entire process of in‑situ formation and annihilation of microscopic Ag filaments under voltage driving, but also rigorously verified that the AgBiS₂ lattice maintains high structural integrity and chemical stability after many complete polarity reversal cycles. This characterization approach, combining high spatial resolution with real‑time electrical manipulation, provides irrefutable experimental evidence for the underlying working mechanism of bipolar optoelectronic synaptic weights.

Overcoming the manufacturing barrier of scaling from a single device to an arrayed chip is the core breakthrough of this research in semiconductor integration processes. The high‑temperature growth conditions of traditional neuromorphic optoelectronic materials often conflict with existing mature silicon‑based processes. In contrast, the AgBiS₂ material used in this study has the excellent characteristic of low‑temperature processing below 200 °C and fully supports high‑precision lithographic patterning at the micrometer scale. Leveraging this manufacturing advantage, the research team successfully performed non‑destructive monolithic integration of thousands of photodiode pixel units with thin‑film transistor readout circuits, fabricating a high‑resolution, broad‑spectrum image sensor chip. This integrated array exhibits extremely low spatial crosstalk and high device electrical uniformity, being fully compatible with standard microelectronics manufacturing platforms. It paves the way for the scalable fabrication of next‑generation, megapixel‑scale in‑sensor computing smart vision chips aimed at industrial mass production.

Based on the developed novel image sensor chip, the research team successfully demonstrated a dual‑mode driven edge intelligent machine vision perception system. In the symmetric mode for non‑destructive information acquisition, the array sensor, with its excellent broad‑spectrum response capability, excelled in performing infrared see‑through imaging tasks under complex occlusion environments. In the asymmetric mode for in‑focus computation, the sensor utilized in‑situ configured non‑volatile photoconductive weights to directly perform optoelectronic convolution operations such as feature extraction, edge detection, and parallel pattern recognition at the optical signal receiving end. The research team further deeply integrated this in‑sensor computing hardware architecture with drone and robotic arm control systems to construct a hybrid convolutional neural network, achieving a human‑machine interaction application for eye‑tracking with extremely high recognition accuracy and nanosecond‑level ultra‑low latency, fully demonstrating the great potential of this technology in AIoT (Artificial Intelligence of Things) terminals.

ZEPTOOLS, as a high‑end precision instrument company based in China, is a leading supplier of in‑situ electron microscopy characterization solutions. Its PicoFemto series of in‑situ transmission electron microscopy characterization solutions have continuously provided technical support for major research achievements of users both domestically and internationally. The figure below shows the ZEPTOOLS in‑situ TEM sample holder product used in this research achievement: