Neuromorphic computing devices based on the asymmetric temperature gradient

Abstract

Neuromorphic computing devices, which emulate biological neural networks, are crucial in realizing artificial intelligence for information processing and decision-making. Different types of neuromorphic computing devices with varying resistance levels have been developed, such as oxide-based memristors caused by ion diffusion, phase transition-based devices caused by threshold switching, progressive crystallization/amorphization, and spintronics-based devices caused by magnetic domain switching. However, these devices face significant challenges, including disruptions in the reading process, limited scalability in integrated circuits, and non-linearity in weight change. To address these challenges, alternative approaches are required. In this study, we introduce a multi-layer-multi-terminal neuromorphic computing device based on the asymmetric temperature gradient. Our device exhibits a wide range of synaptic functions, including potentiation, depression, and both anti-symmetric and symmetric spike-timing-dependent plasticity. The thermal driving strategy offers an energy-efficient platform for future neuromorphic computing devices to achieve artificial intelligence.