Fabrication And Performance Study Of Electrolyte-gated Double-layer Synaptic Transistor

In recent years,due to the traditional von Neumann computers high energy consumption,large size,and other bottlenecks,it has been difficult to meet the needs of future artificial intelligence.It is worth noting,however,that brain-like neuromorphic systems based on neural algorithms exhibit the potential to overcome von Neumann bottlenecks and transcend Moore’s law.In recent scientific research,the ion-conducting electrolyte-gated electric double layer transistor（EDLT）is so short that the electric field can easily exceed 10 MV/cm,compared with other structures of the neuromorphic devices,to achieve better gate control.At the same time,because electrolytes only conduct certain types of ions and insulating electrons,there is only a very low leakage current between the channel material and the control terminal,making it easy to achieve low power consumption.Therefore,double-layer electrolyte-gated transistors（EGT）are promising candidates for synaptic electronics and neuromorphological systems.In this study,a low-cost sol-gel technique was used to fabricate a double-layer synaptic transistor based on two different electrolyte systems,and the synaptic plasticity simulation and neuromorphic computation were realized:（1）Mg-doped SnO₂ double-layer synaptic transistors were fabricated on Si/SiO₂substrates by the sol-gel method using LiClO₄/PEO polymer as the gate electrolyte.It was found that a 5 wt%Mg-doped SnO₂ electrolyte-gated transistor（abbreviated as MgSnO EGT）had the best electrical performance.Using the device to simulate synaptic behavior and compute neural morphology,a variety of synaptic plasticities with both excitation and inhibition patterns have been successfully realized,including single-pulse excitatory/inhibitory post-synaptic current（EPSC/IPSC）,dual-pulse excitatory/inhibitory synaptic plasticity（paired pulse facilitation/depression,PPF/PPD）,spiking-rate-dependent plasticity（SRDP）,and spike-timing-dependent plasticity（STDP）.A"heart"array composed of 25 MgSnO EGTs was constructed to simulate the learning and memory functions of the human brain.After training,the"heart"pattern was successfully stored in the synaptic arrays.In the multi-gate mode,the logic functions"AND","OR","NAND"and"NOR"are simulated.At the same time,the classical Ivan Pavlov is successfully simulated by using the multi-gate structure,and the effects of training pulse width and pulse number on learning efficiency are studied.In neuromorphic computation,an artificial neural network based on MgSnO EGT can achieve 92.3%recognition accuracy for 8×8 pixels of handwritten digital images extracted from the"Optical Recognition of Handwritten Digits"data set.（2）In₂O₃ double-layer synaptic transistors were fabricated on a Si substrate by the sol-gel method using ZrO_x and LiZrO_x solid-state electrolytes as gate electrodes,respectively.The results show that the In₂O₃ double-layer synaptic transistor based on LiZrO_x solid-state electrolyte exhibits better large-lag behavior in the double-sweep transfer curve because of Li⁺doping and also exhibits a stronger postsynaptic current response.Therefore,the device is used to simulate synaptic behaviors,such as typical short-range plasticity and long-range plasticity synaptic behaviors,such as EPSC,PPF,SRDP,STDP,and LTP/LTD（long-term potentiation or depression）.At the same time,the low energy consumption of a single pulse（55.3 f J）is achieved,which is at a low level in the same system of neuromorphic devices.By improving the number of states of electrical conductivity,the dynamic conductance range(G_max/G_min)>100 at 128 states was achieved,which was applied to image recognition.High-precision recognition of near-ideal values is achieved in the classification recognition of file types from the small image version（8×8 pixels）from the"Handwritten digital optical recognition"dataset,the large image version（28×28 pixels）from the MNIST dataset,and the file types from the Sandia File Classification Dataset.By using different voltage sequences to change the channel resistance state,three conductance modes of high,medium,and low conductance are realized,which proves that the EGT can be used as a multilevel memory device.At the same time,the recognition accuracy of the above three types of conductance patterns is compared.In order to apply this device to wireless communication,Morse code in wireless communication is simulated.