以下文章来源于Artificial Synapse ，作者Synapse
研究背景
信息的获取、处理和传输是通过生物突触互连网络之间连接强度的动态变化实现的。它们通过神经递质的交换从轴突末端穿过突触间隙与受体树突进行交流；根据刺激的性质，神经元可以产生不同的神经递质，对树突产生不同的影响；这种适应性是完成一系列复杂功能的基础，可以使我们做出适当的反应，以适应现实世界中不断变化的条件。将感知、学习、计算和记忆等多种功能集成到神经形态设备中已成为一种重要趋势。
纳米线结构的神经形态晶体管正被广泛应用于人工感觉和运动系统。由于纳米线结构具有细长的形状、灵活性和良好的大面积可扩展性，因此适合模拟神经纤维的形态，以实现类似神经的高集成度目标。然而，大多数已报道的 NW 结构突触器件只能模拟单一神经递质的释放，无法实现短期和长期之间的即时可塑性切换。此外，由于单型半导体材料或单通道结构的限制，制造能有效感知外部信息 (电和光刺激)的人工突触一直是一项具有挑战性的任务。受此启发，作者提出了一种特殊的 p-i-n 结，利用电子和光学模式调制载流子注入，进一步扩大感知外部信息的能力。
研究成果
南开大学徐文涛教授团队展示了一种光电神经形态晶体管 (PENT)，它由垂直相分离的 二维C8-BTBT/PMMA混合异质结和连续生长的一维ZnO纳米线(NWs) 阵列组成。这是首次报道在基于纳米线的 PENT 上开发出 p-i-n 混合异质结。该器件集成了光学传感和电子处理功能。在光学传感方面，该器件可对低至每平方厘米微瓦特的低强度紫外线(UV)做出响应，是迄今为止对紫外线最敏感的非易失性光电神经器件。在电处理方面，PENT 通过改变导电通道中的电荷载流子，展示了短期和长期之间的机械即时可切换可塑性；这仿真了同一轴突选择性释放不同的神经递质，即多巴胺和去甲肾上腺素，首次实现了斯金纳箱的避痛和愉悦诱导模式。此外，还开发了一种新的手势识别策略。通过将类脑处理和人工感觉神经相结合，这种方法可用于控制和设计多用途神经形态系统。相关研究以“A low-dimensional hybrid p-i-n heterojunction neuromorphic transistor with ultra-high UV sensitivity and immediate switchable plasticity”为题发表在Applied Materials Today期刊上。
图文导读
Fig. 1. (a) Under electro-optical pulse, a typical PENT composed an two metal contact pads, and Q1D ZnO NWs with a 2D PMMA/C8-BTBT decorated sheath, which emulates the perception of ultraviolet light and the conversion of photoelectric signals; an ion gel as the presynaptic membrane to combine with PENT for electrical signal transmission. (b) A cross-sectional view of a typical PENT as an optical sensing unit. (c) A cross-sectional view of a typical PENT as an electrical processing unit, respectively. (d) Optical microscope (OM) image of highly-aligned ZnO NWs. (e) AFM images with height profiles of PMMA/C8-BTBT. (f) XPS survey spectra of O 1 s of ZnO NWs. (g) XRD patterns of ZnO NWs on SiO2, and of PMMA/C8-BTBT film on ZnO NWs. (h) UV-Vis-NIR absorption spectrum of different films on glass substrate. (i) PL spectra of different films on glass substrate. PL excitation wavelength was 325 nm.
Fig. 2. (a) Mechanism of the PENT for optical response at positive bias. (b) [1] EPSC for PENT triggered by illumination at wavelengths of 380 nm. Incident power density = 7 μw/cm2, VDS = 1 V. (c) Pain-avoiding of UV keratitis for PENT devices with different structures triggered by UV illuminating of various duration. (d) Retention current of PENT after UV illuminating of 11 s. (e) A voltage programmed crossbar array using PENT devices and a mask with Nankai University badge are drawn to illustrate the simulation for optical storage. (f) Weight distribution read after different time for UV illumination of 3.5 s.
Fig. 3. (a) Schematic of the transmissions of two kinds of excitatory neurotransmitters (dopamine and noradrenaline) in the synaptic cleft under pre-synaptic spikes, under different internal environments. (b) PSCs for PENT triggered by different external spikes (spikes = -2 and 4 V, 10 ms), under internal environment of VDS = ± 0.5 V. (c) PPF index triggered by spike = 4 V versus interval time between successive spikes. (d) Current gain for PENT triggered by 30 external spikes of 4 V, under different internal environment (VDS = ± 0.5 V). (e) 30 s after 100 external spikes of 4 V were removed, retention rates of PENTs with different structures, under different internal environment (VDS = ± 0.5 V). (f) The schematic and band diagrams of the junction under zero VDS, reverse VDS, and forward VDS. (g) SDDP index (An/A1 × 100%) triggered by different spikes (spike = 1–4 V), according to spike duration from 10 to 70 ms, under different internal environments (VDS = ± 0.5 V). (h) SFDP index (An/A1 × 100%) triggered by a series of spikes of 4 V, according to the spike frequency from 2.5 to 50 Hz, under different internal environments (VDS = ± 0.5 V). (i) SNDP index (An/A1 × 100%) triggered by a series of spikes of 4 V with different number from 1 to 30, under different internal environments (VDS = ± 0.1, ± 0.2, ± 0.3, ± 0.4 and ± 0.5 V).
Fig. 4. (a) Analog weights update under different internal environments (VDS = ± 0.5 V). The continuous increase of channel conductance G caused by a series of repeated positive spikes of 4 V, and the decrease of G triggered by a series of repeated negative spikes of -2 V. (b) ANN for gesture recognition (6 types of gestures) through the full connection of 104 synaptic weights. (c) The recognition rate of gesture recognition as a function of the training epoch according to the two cases. (d) Confusion matrix of extraction results according to the two cases. (e) Weight values of 10,000 input synapses during 2000 learning phases, under two cases. (f,g) Schematic and implementation of Skinner Box using PENT (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.).
总结与展望
作者将 一维 ZnO NWs 与二维 PMMA/C8-BTBT 相分离薄膜相结合，制造出了一种具有p-i-n 混合异质结结构的电子光学神经形态晶体管。这种多维异质结集成了光学传感和电子处理功能。这种异质结进一步提高了 PENT 的光学灵敏度。实验表明，该器件能以超低紫外光强度感测光产生的图案，并记忆图像。通过改变电荷载流子的传输方向，它首次实现了多巴胺主导的 STP 和去甲肾上腺素主导的 LTP 之间可立即切换的突触可塑性。多路复用神经传递过程的成功模拟可以帮助 PENT 模拟经典的操作性条件反射。此外还实现了突触权重的宽动态工作范围这一特性有望应用于手势识别。这项工作为传感记忆神经形态电子学的控制和设计提供了一种潜在的方法。
文献链接
A low-dimensional hybrid p-i-n heterojunction neuromorphic transistor with ultra-high UV sensitivity and immediate switchable plasticity
https://doi.org/10.1016/j.apmt.2021.101223
转自：“i学术i科研”微信公众号
如有侵权，请联系本站删除！
