State number,operation power,dynamic range and conductance weight update linearity are key synaptic device performance metrics for high-accuracy and low-power-consumption neuromorphic com-puting in hardware.However,hi...State number,operation power,dynamic range and conductance weight update linearity are key synaptic device performance metrics for high-accuracy and low-power-consumption neuromorphic com-puting in hardware.However,high linearity and low power consump-tion couldn’t be simultaneously achieved by most of the reported synaptic devices,which limits the performance of the hardware.This work demonstrates van der Waals(vdW)stacked ferroelectric field-effect transistors(FeFET)with single-crystalline ferroelectric nanoflakes.Ferroelectrics are of fine vdW interface and partial polar-ization switching of multi-domains under electric field pulses,which makes the FeFETs exhibit multi-state memory characteristics and ex-cellent synaptic plasticity.They also exhibit a desired linear conduc-tance weight update with 128 conductance states,a sufficiently high dynamic range of G_(max)/G_(min)>120,and a low power consumption of 10 fJ/spike using identical pulses.Based on such an all-round device,a two-layer artificial neural network was built to conduct Modified Na-tional Institute of Standards and Technology(MNIST)digital num-bers and electrocardiogram(ECG)pattern-recognition simulations,with the high accuracies reaching 97.6%and 92.4%,respectively.The remarkable performance demonstrates that vdW-FeFET is of obvious advantages in high-precision neuromorphic computing applications.展开更多
文摘State number,operation power,dynamic range and conductance weight update linearity are key synaptic device performance metrics for high-accuracy and low-power-consumption neuromorphic com-puting in hardware.However,high linearity and low power consump-tion couldn’t be simultaneously achieved by most of the reported synaptic devices,which limits the performance of the hardware.This work demonstrates van der Waals(vdW)stacked ferroelectric field-effect transistors(FeFET)with single-crystalline ferroelectric nanoflakes.Ferroelectrics are of fine vdW interface and partial polar-ization switching of multi-domains under electric field pulses,which makes the FeFETs exhibit multi-state memory characteristics and ex-cellent synaptic plasticity.They also exhibit a desired linear conduc-tance weight update with 128 conductance states,a sufficiently high dynamic range of G_(max)/G_(min)>120,and a low power consumption of 10 fJ/spike using identical pulses.Based on such an all-round device,a two-layer artificial neural network was built to conduct Modified Na-tional Institute of Standards and Technology(MNIST)digital num-bers and electrocardiogram(ECG)pattern-recognition simulations,with the high accuracies reaching 97.6%and 92.4%,respectively.The remarkable performance demonstrates that vdW-FeFET is of obvious advantages in high-precision neuromorphic computing applications.
