Neuromorphic Devices based on van der Waals Materials for Power-Efficient Computing

The conventional von Neumann computer architecture faces limitations in speed and scalability for data-intensive AI applications due to the physical separation of memory and processing units. Meanwhile, the high power consumption of AI models has led to concerns about global energy and environmental costs. To address these challenges, specialized computing architectures and hardware based on new principles are demanded. Neuromorphic computing architectures, inspired by human brain’s structure and function, offer a promising solution by enabling parallel, event-driven and in-memory computing. The unconventional computing architecture demands an exploration of unconventional electronic materials and device structures. The two-dimensional (2D) materials, with van der Waals nature, offer a particularly compelling platform for neuromorphic computing due to their unique properties, atomically sharp interface, ultralow lightweight, scalability and compatibility with CMOS technology.

This dissertation employs a combined experimental and computational approach to systematically investigate MoS₂-based synaptic devices and demonstrate their potential for power-efficient machine learning. In particular, we focus on building a physical reservoir computing system based on synaptic transistors as virtual nodes using MoS2 as the channel material. At the device level, we experimentally demonstrate a binary oxide based three terminal MoS2 synaptic device with a constantly low (~50 pW) programming power consumption across a wide dynamic range of 105. Additionally, we develop a physics-based device simulator to capture the underlying device physics that governs the characteristics of the synaptic device. At the system level, we build a time-delay physical reservoir utilizing the intrinsic nonlinearity and dynamics of the synaptic device, and achieved > 92% accuracy in the benchmarked MNIST hand-written digit classification test as a demonstration. This work opens up new avenues by leveraging the unique properties of van der Waals 2D materials to build novel and scalable neuromorphic computing systems as power efficient AI hardware.