Investigating ferroelectric and metal-insulator phase transition devices for neuromorphic computing
MetadataShow full item record
Neuromorphic computing has been proposed to accelerate the computation for deep neural networks (DNNs). The objective of this thesis work is to investigate the ferroelectric and metal-insulator phase transition devices for neuromorphic computing. This thesis proposed and experimentally demonstrated the drain erase scheme in FeFET to enable the individual cell program/erase/inhibition for in-situ training in 3D NAND-like FeFET array. To achieve multi-level states for analog in-memory computing, the ferroelectric thin film needs to be partially switched. This thesis identified a new challenge of ferroelectric partial switching, namely “history effect” in minor loop dynamics. The experimental characterization of both FeCap and FeFET validated the history effect, suggesting that the intermediate states programming condition depends on the prior states that the device has gone through. A phase-field model was constructed to understand the origin. Such history effect was then modelled into the FeFET based neural network simulation and analyze its negative impact on the training accuracy and then propose a possible mitigation strategy. Apart from using FeFET as synaptic devices, using metal-insulator phase transition device, as neuron was also explored experimentally. A NbOx metal-insulator phase transition threshold switch was integrated at the edge of the crossbar array as an oscillation neuron. One promising application for FeFET+NbOx neuromorphic system is to implement quantum error correction (QEC) circuitry at 4K. Cryo-NeuroSim, a device-to-system modeling framework that calibrates data at cryogenic temperature was developed to benchmark the performance of the FeFET+NbOx neuromorphic system.