Artificial Synapses Based On Ferroelectric Tunnel Functions And Their Application In Neuromorphic Computing

The separation of memory and processor in traditional Von-Neumann architecture computer results in significant time and energy consumption for data scheduling,known as the"bandwidth bottleneck".The best way to break through this bottleneck is to build a brain-like neuromorphic computing system that integrates storage and computing,and the basic components of the brain-like computing system are artificial neurons and artificial synapses,so the design and preparation of high-performance artificial neurons and synapses is the basis for building a brain-like neuromorphic computing system.The conductance of memristor can vary with input stimuli,a property similar to biological synaptic plasticity,so memristors are considered strong candidates for artificial synapses.Based on ferroelectric tunnel junction memristor devices,this paper simulates the important neural functions of synapses and applies them to neuromorphic computing such as supervised learning and unsupervised learning.The main research contents and results are as follows:(1)Realize synaptic weight updating with high linearity and high symmetry.By introducing appropriate amount of oxygen vacancies into the BaTiO3 ferroelectric layer,ferroelectric tunnel junctions combining ferroelectric polarization switching and oxygen vacancy migration mechanisms were prepared.Based on the artificial synapse of the ferroelectric tunnel junction,the weight updating with high linearity and high symmetry is realized,and the nonlinearity(NL)is as low as 0.13.The artificial synapse of the ferroelectric tunnel junction is applied to artificial neural network for supervised learning,and its pattern classification recognition accuracy reaches 96.7%,which is the closest two-terminal device to 98%of the theoretical value so far.Even considering the influence of cycling-to-cycling and device-to-device changes on the back-propagation algorithm,the ferroelectric tunnel junction artificial synapse still performs well in supervised learning.In addition,through weight updating with high symmetry,we also achieve spike-time dependent plasticity(STDP)with long-term potentiation(LTP)and long-term depression(LTD)balance,and unsupervised learning based on this artificial synapse has high robustness to noise.The results demonstrate the potential of ferroelectric tunnel junction memristor incorporating oxygen vacancy migration in high-performance neuromorphic computing.(2)Prepare a second-order memristor to realize high-order synaptic function.By epitaxial growth of BaTiO3 ferroelectric layer containing a large number of oxygen vacancies under low oxygen pressure,a second-order memristor device in ferroelectric tunnel junctions was prepared,and the first-order variables were verified to be ferroelectric polarization state and the distribution of oxygen vacancies,and the second-order variable is the concentration gradient of oxygen vacancies.The conductance of the device is affected by the pulse amplitude,pulse width,pulse interval,and pulse number,which can not only realize basic synaptic functions such as paired-pulse facilitation(PPF),spike-rate-dependent plasticity(SRDP),and experiential learning,but also simulate high-order synaptic plasticity such as synaptic metaplasticity and stimulus history-dependent plasticity.The results of this study provide a method for simulating biological synapses in ferroelectric second-order memristors,and provide a device basis for realizing high-order spatiotemporal recognition functions in artificial neural network.