Adaptive Oxide Devices For Brain-Inspired Electronics: From Physics to Deep Learning

The pervasive usage of artificial intelligence to improve the quality of life has led to a massive demand for energy. One of the primary reasons for the tremendous energy consumption in traditional von Neumann architecture is the continuous data transfer between the memory and the processing unit. Brain-inspired analog and in-memory computing aim to solve this issue by allowing calculation and memory at the same place, similar to nature’s astounding computing machine: the human brain. However, the widespread adoption of brain-inspired devices is prohibited by their non-ideal behavior arising from the need for more material, device, and system-level optimization. This dissertation aims to develop a deeper understanding of brain-inspired devices and take strides toward their ideal behavior to unlock their full potential. The first aim of this dissertation focuses on developing a fundamental understanding of the HfOx synaptic device physics and the impact of doping on its characteristics. The synaptic devices exhibit non-volatile resistance change by forming an oxygen vacancy-rich conductive filament (CF) and moving oxygen ions within it. The resistance change is expected to be impacted by changing the number of vacancies through the addition of titanium dopants. The electrical characterization of the fabricated titanium-doped HfOx devices shows that an increase in titanium results in decreased forming voltage and allowable stop voltages during switching. In addition, analysis of the analog switching behavior reveals that the titanium-doped CF, not the bulk oxide, solely governs the analog response. The physical mechanisms responsible for the observed responses are proposed. Additionally, C-STAO, a compact model combining trap-assisted tunneling and ohmic transport was developed with the goal of understanding the device mechanism. Experimental data fitting with C-STAO and COMSOL Multiphysics® simulation results show that the local rupture of the CF occurs near the reset anode. Using this knowledge, the second aim of the dissertation aspires to improve the synaptic devices through barrier layer, electrode, and oxygen reservoir layer optimization and corroborate the proposed hypotheses using simulation. One of the significant challenges for brain-inspired technology is an abrupt resistance change in response to a positive bias called abrupt set, which happens because of the abrupt oxygen ion motion. A hypothesis is proposed that a ~1 nm SiOx barrier layer, with a high oxygen ion migration barrier close to the reset anode, will provide better control over the ion motion. Thus, HfOx/SiOx devices were fabricated, which show a gradual resistance change during set and less variability. In addition a COMSOL Multiphysics® simulation was performed to validate the experimental data. This work also demonstrates that slowing the motion of oxygen ions results in a fundamental trade-off between a gradual set and a low on-off ratio. The low on-off ratio is associated with a high off-current. Further, it is hypothesized that an electrode with low thermal conductivity would reduce heat removal and increase the number of oxygen ions closing the CF, resulting in less off-current. However, the challenge is that most of the electrodes have high thermal conductivity. Ti2AlN, like many other MAX phase materials, has metal-like electrical conductivity and ceramic-like low thermal conductivity. So, HfOx devices with MAX phase as the bottom electrode were fabricated, demonstrating ultra-low off-current, large switching window, high endurance, and multi-level capability. The third aim builds on the insight gained from the device-level research to improve the system-level performance of an IBM analog brain-inspired chip for deep learning. In collaboration with IBM T. J. Watson lab, this aim focuses on optimizing an algorithm for training a deep learning model using a 14 nm technology-based analog accelerator array. The performance of analog AI hardware degrades after repeated usage. An electrical biasing technique, Recovery Stabilization (ReSta) is experimentally demonstrated, which can recover the accuracy of the analog AI hardware up to the level before the introduction of fatigue. Overall, the dissertation's material, device, and system-level investigation advance our understanding of brain-inspired computing for sustainable energy-efficient artificial intelligence.

Date

2024-02-28

Resource Type

Text

Resource Subtype

Dissertation

Full item page