| dc.description.abstract | Energy-intensive artificial intelligence (AI) is prevailing and changing the world, which requires energy-efficient computing technology. However, traditional AI driven by von Neumann computing systems suffers from the penalties of high-energy consumption and time delay due to frequent data shuttling. To tackle the issue, brain-inspired neuromorphic computing that performs data processing in memory is developed, reducing energy consumption and processing time. Particularly, some advanced neuromorphic systems perceive environmental variations and internalize sensory signals for localized in-senor computing. This methodology can further improve data processing efficiency and develop multifunctional AI products. Memristive devices are one of the promising candidates for neuromorphic systems due to their non-volatility, small size, fast speed, low-energy consumption, etc.
In this thesis, memristive devices based on chalcogenide and metal-oxide materials are fabricated for neuromorphic computing systems. Firstly, a versatile memristive device (Ag/CuInSe2/Mo) is demonstrated based on filamentary switching. Non-volatile and volatile features are coexistent, which play multiple roles of non-volatile memory, selectors, artificial neurons, and artificial synapses. The conductive filaments’ lifetime was controlled to present both volatile and non-volatile behaviours. Secondly, the sensing functions (temperature and humidity) are explored based on Ag conductive filaments. An intelligent matter (Ag/Cu(In, Ga)Se2/Mo) endowing reconfigurable temperature and humidity sensations is developed for sensory neuromorphic systems. The device reversibly switches between two states with differentiable semiconductive and metallic features, demonstrating different responses to temperature and humidity variations. Integrated devices can be employed for intelligent electronic skin and in-sensor computing. Thirdly, the memristive-based sensing function of light was investigated. An optoelectronic synapse (ITO/ZnO/MoO3/Mo) enabling multi-spectrum sensitivity for machine vision systems is developed. For the first time, this optoelectronic synapse is practical for front-end retinomorphic image sensing, convolution processing, and back-end neuromorphic computing. This thesis will benefit the development of advanced neuromorphic systems pushing forward AI technology. | en |
