神经形态工程学
计算机科学
接口(物质)
预处理器
存水弯(水管)
能源消耗
人工智能
光电子学
材料科学
过程(计算)
电子工程
人工神经网络
能量(信号处理)
人类视觉系统模型
动态随机存取存储器
突触
深度学习
图像传感器
电介质
计算机硬件
嵌入式系统
光开关
电容
高效能源利用
图像压缩
晶体管
可视化
作者
Xin Qun Feng,Zhiyu Bao,Ziyang Shi,Zeyang Ren,Kai Su,Yachao Zhang,Weihang Zhang,Yanjing He,Zhihong Liu,Yue Hao,Jincheng Zhang
出处
期刊:InfoMat
[Wiley]
日期:2026-11-13
卷期号:: e70203-
摘要
The popular artificial intelligence era has promoted the development of energy‐efficient artificial visual systems which require optoelectronic synapses capable of co‐integrating sensing, memory, and computation. Here a wafer‐scale GaN HEMT with gate‐dielectric engineering is investigated as a hierarchical neuromorphic vision demonstration platform integrating sensing, memory, and computing. It is found that the HfO 2 ‐gated GaN HEMT has oxygen‐vacancy‐related trap states in the dielectric layer and higher interface trap states with deeper energy levels in interface trap states, which can significantly enhance the trapping and slow release of photogenerated carriers, thereby enhancing the synaptic performance. As a result, this device achieves balanced and competitive performance metrics with a PPF of 79.34%, conductance‐update linearity of 0.81 and single‐pulse energy consumption of 1.11 μJ/event. In application, it completes system‐level validation from in‐sensor spatiotemporal compression with a 79.12% accuracy for dynamic trajectory recognition to FCNN/physical RC readout achieving static handwritten‐digit recognition with a high accuracy of 97.69%, demonstrating the possibility of using the same hardware platform to process static images and dynamic temporal information. This work not only provides experimental evidence for understanding the carrier trapping/release dynamics, builds a reproducible device‐engineering pathway for performance optimization of GaN‐based optoelectronic synapses, but also promotes the development of artificial visual hardware from biological visual preprocessing to advanced neural‐network computation. image
科研通智能强力驱动
Strongly Powered by AbleSci AI