Key Laboratory of Photoelectronic Thin Film Devices and Technology of Tianjin, Institute of Photoelectronic Thin Film Devices and Technology of Nankai University, Tianjin 300350, China
Fund Project:Project supported by the Key Area R&D Program of Guangdong Province, China (Grant No. 2018B030338001), the Tianjin Science Foundation for Distinguished Young Scholars, China (Grant No. 19JCJQJC61000), the Natural Science Foundation of Tianjin, China (Grant No. 18JCYBJC16000), and the Fundamental Research Funds for the Central Universities, China (Grant Nos. 075-63191740, 075-63191745)
Received Date:16 June 2020
Accepted Date:26 October 2020
Available Online:30 November 2020
Published Online:05 December 2020
Abstract:Recently, neuromorphic systems capable of parallel information processing have attracted increasing attention. A neuromorphic system is desired to emulate a human brain, which consists of hundreds of billions of neurons connected with even more synapses. Synapses are important connection parts between neurons to transmit information through release and reception of neurotransmitters. A neuromorphic system could replicate brain learning, cognition and computation of a human brain to process huge data with 1016 floating point numbers per second. The high computing efficiency has attracted many researchers to study artificial synapses for application in future artificial intelligence. The synaptic weight could be adjusted by the received information. This provides a basis for the learning and computing capability of artificial synapses. So far, a number of semiconductor materials have been used in artificial synaptic devices, like some organic materials, e.g. Poly(3-hexylthiophene-2,5-diyl)(P3HT), [1]Benzothieno[3,2-b][1]benzothiophene, 2,7-dioctyl-(C8-BTBT) etc, some inorganic oxides such as zinc oxide, indium zinc oxide(IZO), indium gallium zinc oxide(IGZO), transition metal oxides, etc, and two-dimensional materials, e.g. graphene, black phosphorus, and organic-inorganic hybrid perovskite materials. Among them, transition metal oxides are attractive due to their unique layered structure and inherent properties, which are important in photohydrolysis, lithium ion batteries, and field-effect transistors. MoO3, as a typical transition-metal oxide, has been used in artificial synaptic devices, with different preparation methods, such as mechanical exfoliation, chemical vapor deposition (CVD) and chemical vapor transportation (CVT), pulse-laser deposition (PLD). Here, we report the preparation of a semiconductor layer of MoO3 nanosheets by hydrothermal method, and the use of a TiO2 nanoparticle seed layer to improve the adhesion of MoO3 nanosheets. This is a cost-effective and controllable process. The high surface-to-volume ratio of the material provides large contact area at the interface to allow easy ion diffusion. The device emulates important synaptic functions, such as excitatory post-synaptic current (EPSC), paired-pulse facilitation (PPF), spike-duration dependent plasticity (SDDP), spike-voltage dependent plasticity (SVDP) and spike-rate dependent plasticity (SRDP). This work could be an important addition to the neuromorphic research field. Keywords:artificial synapse/ hydrothermal synthesis/ molybdenum trioxide nanosheets/ synaptic plasticity
突触可塑性是指突触连接强度可以被持续调节的性质, 会随着自身活动的增强或减弱相应的增强或减弱[37]. 突触可塑性可分为两种: 长程可塑性(LTP)和短程可塑性(STP). STP往往可以实现短时程行为、短时程记忆的重要功能, 可以在大脑接收大量刺激信号时过滤掉不重要的信息, 是提取有价值的信息的关键[38]. EPSC是突触后膜收到刺激后短暂去极化所形成的现象, 是STP的重要表现之一. 为了在制备的人工突触电子器件上模拟这种突触兴奋性行为, 以顶电极和底电极作为突触前膜和突触后膜, MoO3活性层模拟突触间隙. 在底电极上施加一个恒定的偏置电压(0.01 V)作为读取电压, 将幅值为1 V的电压施加在顶电极上建立一个临时的外加电场. 外加电压导致了EPSC的产生, 器件受到刺激时电流瞬时增加, 在刺激消失之后突触后电流快速衰退恢复至初始状态如图3(a)所示. 这一现象的产生可能是在器件受到刺激的瞬间, 固体电解质中的Li+在外加电场的作用下有序迁移, 聚集在MoO3表面, 引起MoO3中电子的定向移动从而造成突触后电流的瞬时增加. 消除刺激之后电荷扩散到其平衡位置, 因此电流衰减恢复至初始状态. 这一现象与生物性神经突触中的EPSC现象对应, 一旦适当的刺激到达轴突末端并激活Ca2+通道, 从突触小泡释放的神经递质会迅速扩散穿过突触间隙, 并与突触后细胞膜上的这些受体结合, 最终产生兴奋性突触后电流[39]. 图 3 (a) 单个幅值为1 V的刺激在MoO3人工神经突触上引起的EPSC; (b)一对幅值为1 V的刺激在MoO3人工神经突触上引起的PPF; (c)和(d) 多对时间间隔不同, 幅值为1 V的脉冲引起的PPF及PPF Index Figure3. (a) EPSC triggered by a single 1 V spike at a MoO3 artificial synapse; (b) PPF triggered by a pair of 1 V spikes at a MoO3 artificial synapse; (c) and (d) PPF and PPF index triggered by spikes with different time intervals and same amplitudes of 1 V.
PPF也是突触短时可塑性的一个重要表现, 在视觉和听觉系统中接收和解读瞬时消息中至关重要[40], PPF体现的是在前突触上施加两次连续的刺激, 第二次刺激造成的突触后兴奋电流高于前一次刺激的现象, 在此器件上模拟的PPF现象如图3(b)所示. 相同幅值电压下的两次刺激, 第二次刺激造成的EPSC与第一次的相比增加17.94%. 在第一次刺激下, Li+的迁移导致MoO3活性层中电子的定向移动, 刺激消失后离子逐渐迁移回到其平衡位置. 但如果第二次刺激在其到达平衡位置之前到达, 会有更多数量的离子在临时电场的作用下发生定向移动从而导致活性层中更多的电子发生定向移动, 所以第二次EPSC高于第一次. 并且高出值的大小由两次之间的时间间隔有关, 越小的时间间隔会造成更大的EPSC[41]. 相同振幅的电压刺激下, 不同时间间隔的PPF如图3(c)所示, 可以很明显的看出, 随着时间间隔的增加第二次EPSC增益逐渐减小, 在这里将增益PPF Index定义为(A2 – A1)/A1, 其中A1是第一次刺激后的EPSC峰值, A2是连续两次刺激后的EPSC峰值. 持续重复的刺激可以使得突触之间的连接性和信息传递性增加, 联系更加紧密[42]. 在神经突触学习规则当中, SRDP描述的是突触前神经元接收不同频率刺激后释放不同数量的神经递质, 从而导致突触后神经元产生的信号电流不同的现象. 在此器件上施加0.01 V的读取电压, 振幅为1 V的刺激电压, 以不同的时间间隔来调制施加刺激的频率, 结果如图4(a)所示. 从实验结果可以看出, 随着施加的刺激频率的降低, 突触后兴奋电流增长的趋势逐渐降低最后趋于平缓. 这一趋势的产生是由于在较高频率的刺激下, 最初产生EPSC的离子在没有回到其平衡位置时下一次刺激就已经到达所以产生了与PPF类似的趋势. 而当施加的刺激频率较低时, 在下一次刺激达到之前, 造成突触后电流的离子大部分已经回到其平衡位置, 所以形成的趋势较为平缓. 图 4 (a) 施加幅值为1 V刺激个数分别为1, 2, 3, 5, 8在MoO3突触器件上引起的的SRDP; (b) 相同幅值不同刺激持续时间造成的SDDP; (c) 幅值分别为0.5, 1.0, 1.5, 2.0 V的刺激在突触器件上引起的SVDP; (d)为施加幅值为0.2 V时所获得的兴奋性突触后电流及完成一次信号传递所消耗的能量 Figure4. (a) SRDP on MoO3 synapses triggered by the number of spikes of 1 V applied at 1, 2, 3, 5, and 8; (b) SDDP triggered by different spike duration time with the same amplitude; (c) SVDP triggered by spikes with amplitudes of 0.5, 1.0, 1.5 V, and 2.0 V on synaptic devices; (d) the excitatory postsynaptic current when the applied spike is 0.2 V and the energy consumed to complete a signal transmission.