Ultra-highly linear Ga2O3-based cascade heterojunctions optoelectronic synapse with thousands of conductance states for neuromorphic visual system
image: Figure | Linear optical potentiation/electrical depression behaviors and high-accuracy image classification functions. a, Eight cycles of optical potentiation/electrical depression behaviors. b, Typical cycle of optical potentiation/electrical depression behaviors under consecutive training pulses, presenting a high linearity of weight updating. c, Schematic illustration of NVS for the application of fingerprint classification. d, The fingerprint images with different-level Gaussian noise, whose standard deviation (std) ranges from 0 to 0.15. e, Recognition accuracy evolution with training epochs for the NVS with different noise ratios. The classification output results depicted in the confusion matrixes f, without and g with Gaussian noise (std is set as 0.1).
Credit: Ya Lin et al.
The rapid development of artificial intelligence has created an urgent need for vision systems that can replicate the human brain's seamless integration of sensing, memory, and processing. While visible-light neuromorphic systems have seen significant progress, deep ultraviolet (DUV) vision remains constrained by fundamental material limitations. DUV light's unique advantages - including minimal atmospheric interference and exceptional sensitivity to organic residues - make it ideal for critical applications ranging from missile defense to biochemical detection. However, conventional DUV optoelectronic synapses suffer from rapid carrier recombination that severely restricts their conductance states and computational precision, creating a technological bottleneck for advanced neuromorphic applications.
In a groundbreaking study published in Light: Science & Applications, researchers from Northeast Normal University have developed a transformative solution using Ga₂O₃-based cascade heterojunctions. Led by Professors Haiyang Xu, Zhongqinag Wang and Ya Lin, the team created an optoelectronic synapse that achieves three landmark breakthroughs: an ultrahigh responsivity of 65,000 A/W (300 times greater than conventional DUV sensors), an unprecedented 4,096 conductance states, and near-perfect linearity with a fitting coefficient of 0.992. These achievements overcome longstanding challenges in carrier recombination and nonlinear weight updates that have limited neuromorphic DUV systems.
The device's exceptional performance stems from its ingeniously designed GTO/Al/HfOₓ heterostructure. At the GTO/HfOₓ interface, a built-in electric field efficiently separates photogenerated electron-hole pairs under DUV illumination. Simultaneously, the Al/HfOₓ Schottky barrier prevents hole collection at the electrode, instead trapping them in oxygen vacancies within the HfOₓ layer. This dual mechanism creates a gain effect that amplifies photocurrent while enabling precise control over thousands of distinct conductance states - a crucial requirement for neuromorphic computing.
The technology's capabilities were demonstrated through several groundbreaking applications. The synapses successfully executed logical AND/OR operations with remarkable fault tolerance, maintaining functionality even with 40% input signal fluctuations. More impressively, the system performed in-sensor arithmetic operations including addition (18+9) and multiplication (9×3) using purely optical inputs, bypassing the von Neumann bottleneck that plagues conventional computing architectures. For real-world validation, the team implemented a fingerprint recognition system that achieved 99.6% classification accuracy - a performance that remained robust (94.7% accuracy) even when subjected to significant noise interference.
"This work represents a paradigm shift in DUV neuromorphic technology," explained Prof. Lin. "By solving both the carrier recombination problem and nonlinear weight update challenge simultaneously, we've created a platform that bridges the gap between biological vision and artificial systems." Prof. Xu added: "The military and security applications are particularly compelling, but we're equally excited about potential uses in assistive technologies and environmental monitoring where DUV's unique properties offer distinct advantages."
Looking ahead, the researchers emphasize that the technology's compatibility with standard semiconductor manufacturing processes positions it for rapid translation to practical systems. Potential near-term applications include real-time surveillance systems that leverage DUV's noise-free operation, and edge AI devices requiring ultra-low power consumption. The team is currently collaborating with industrial partners to scale up production and explore integration with existing sensor networks, potentially ushering in a new era of brain-inspired computing for specialized DUV applications.
The Internet of Things (IoT) and cyber physical systems have opened up possibilities for smart cities and smart homes, and are changing the way for people to live. In this smart era, it is increasingly demanded to remotely monitor people in daily life using radio-frequency probe signals. However, the conventional sensing systems can hardly be deployed in real-world settings since they typically require objects to either deliberately cooperate or carry an active wireless device or identification tag. Additionally, the existing sensing systems are not adaptive or programmable to specific tasks. Hence, they are far from efficient in many points of view, from time to energy consumptions.