Optimization And Application Of Artificial Neural Networks For Hafnium-based Ferroelectric Memory

In the context of the Fourth Industrial Revolution,efficient storage,transmission,and processing of massive data has become one of the major challenges in the field of electronic information.In-memory computing rely on the development of artificial neural networks and emerging non-volatile memory,reducing the information transfer between memory and processors and eliminating the mismatch between storage and processor performance,potentially breaking the bottleneck caused by the von Neumann architecture.This article focuses on the artificial neural network of the Hf0.5Zr0.5O2(HZO)-based ferroelectric memory cells,conducts design optimization and application verification.At the device level,the preparation process of ferroelectric capacitor devices is introduced,and the bionic function with biological synaptic properties is realized through pulse scheme design and high-precision electrical testing;At the array level,an optimized design method for a crossbar array is proposed to compensate for the equivalent resistance caused by interconnect resistance effects.At the network level,a full ferroelectric field-effect transistor(FeFET)reservoir computing network is designed,which achieves classification of target images and prediction of temporal signals.The first part focuses on the preparation and characterization of the hafnium-based ferroelectric memory cells and proposes an optimized design method for a crossbar array structure.Using processes such as magnetron sputtering,atomic layer deposition,and rapid thermal annealing,a TiN/HZO/TiN structure ferroelectric capacitor device array is prepared,and a biomimetic result with biological neural synapse properties is obtained by gradually and accurately controlling the residual polarization state.In the crossbar array,the memory cell is affected by interconnect resistance effects,which limit the further increase of the array size.This article proposes an optimized design method for a crossbar array structure,which compensates for the equivalent resistance change caused by interconnect resistance effects by adding circuits to replace the process of solving Kirchhoff’s equations,thus restoring the calculation accuracy.The second part mainly designs a full FeFET reservoir computing network,which includes a mask layer,a reservoir layer,and a readout layer.On the one hand,based on the nonlinear delay feedback caused by the charge capture phenomenon of FeFET interface defects,a virtual node with short-term memory effect is realized,which is conducive to reducing the number of devices required for the mask layer and the reservoir layer.On the other hand,based on the nonvolatility of the polarization state of ferroelectric materials,a weight iteration with long-term memory effect is realized,which is conducive to the long-term maintenance of the training results of the readout layer.Furthermore,the operation time,connection method,and pulse waveform of the full FeFET reservoir computing network are co-optimized,which is of great significance for the development of neural networks with high compactness,fast speed,and full ferroelectric devices.The third part mainly demonstrates the recognition and prediction functions of the full FeFET reservoir computing network.By classifying and recognizing images from multiple datasets,the network performance of the full FeFET reservoir computing network is evaluated under high compactness and high speed(44 devices,100 ns operation speed).By processing temporal signals,it is confirmed that the full FeFET reservoir computing network has the advantages of low power consumption and fast speed in random signal prediction.By extending the RTN signal,a method for extracting time parameters is provided.This article deeply studies the design optimization and verification application of the hafnium-based ferroelectric storage unit in the artificial neural network.The exploration of the hafnium-based ferroelectric device preparation process is carried out,and the electrical testing results of biomimetic neural synapses are obtained.An improved crossbar array architecture is proposed to compensate for the equivalent resistance change caused by interconnect resistance effects.A full FeFET reservoir computing network is designed,which achieves image classification and prediction of temporal signals.The research content and results of this article are of great significance for the development of energy-efficient,fast,and high-density artificial neural network applications using hafnium-based ferroelectric storage units.