Journal of Materials Chemistry C|基于多端突触记忆晶体管模拟竞争与合作行为 ...

文章链接：https://doi.org/10.1039/D0TC01076F

 APTES记忆晶体管

背景

如今，随着信息产业的发展，基于图灵模型的传统数字计算机面临着“冯·诺依曼瓶颈”。当晶体管的大小达到摩尔定律的极限时，内存和逻辑单元之间的物理隔离导致能量效率逐渐达到饱和。相反，人脑中的生物突触可以并行地执行时空学习和推理，能够有效地克服“冯·诺依曼瓶颈”。人脑包含约10^14个突触，这些突触形成一个复杂的神经网络，当人脑处理大量信息时，突触相互作用非常重要。突触相互作用包括突触竞争。通常，通过在生物系统中扩散可塑性相关蛋白（PRP）来实现突触相互作用。当两个或多个突触之间存在竞争关系时，PRP会流到不同的突触。当发生突触协作时，PRP将流到相同区域，从而增强了发射信号。

摘要

并行模式下的时空学习和推理是神经形态计算（模仿人脑结构）的主要优势。在这里，制造了一个带有自组装层的3-氨丙基三乙氧基硅烷（APTES）的记忆晶体管，以模仿突触行为。在自组装过程中，由于APTES分子之间的氢键而形成的聚集体可能会在界面中形成缺陷。在薄膜晶体管中，不仅可以通过栅电极而且可以通过漏电极有效地捕获空穴，从而在沟道中控制载流子浓度。通过向漏电极施加工作电压，可以发现明显的磁滞窗口。为了研究竞争和合作行为，准备了一个多端膜晶体管，并且可以成功地模仿这些突触行为。最后，仿真了一种基于权重更新法的人工神经网络进行数字判别，总精度达到96.42％。结果表明，基于APTES的薄膜晶体管在突触电子学中具有潜在的应用。

Spatiotemporal learning and inference in parallel mode is the key advantage for neuromorphic computing, which mimics the human brain architecture. Here, a memtransistor was fabricated with a self-assembled layer of 3-aminopropyltriethoxysilane (APTES) to mimic the synaptic behavior. The aggregates due to the hydrogen bonding between APTES molecules during the self-assembled process could form a defect in the interface. In the memtransistor, the holes could be captured effectively through not only the gate electrode, but also the drain electrode, resulting in controlled carrier concentration in the channel. An obvious hysteresis window was found by applying the operation voltage to the drain electrode. To study the competitive and the cooperative behavior, a multi-terminal memtransistor was prepared, and these synaptic behaviors could be successfully imitated. Finally, an artificial neural network based on the weight updating method was simulated to discriminate the numbers, and the total accuracy reached 96.42%. The results demonstrated that the APTES-based memtransistor had potential applications in synaptic electronics.