Wafer-Scale 2D Hafnium Diselenide Based Memristor Crossbar Array for Energy-Efficient Neural Network Hardware

Area of Science:

• Materials Science
• Nanotechnology
• Computer Engineering

Background:

• Neural networks face energy and speed limitations in image processing.
• Implementing crossbar arrays with 2D materials is challenging due to synthesis and integration issues.

Purpose of the Study:

• To develop a memristor crossbar array (CBA) using ultrathin 2D materials for efficient neural network hardware.
• To address challenges in large-scale material synthesis and device integration for CBA implementation.

Main Methods:

• Wafer-scale polycrystalline hafnium diselenide (HfSe2) was grown using molecular beam epitaxy.
• A metal-assisted van der Waals transfer technique was employed for device fabrication.
• Memristor characteristics, synaptic plasticity emulation, and hardware multiply-and-accumulate (MAC) operations were evaluated.

Main Results:

• The memristor exhibited low switching voltage (0.6 V) and energy (0.82 pJ).
• The CBA achieved high artificial neural network recognition accuracy (93.34%) and efficient MAC operations (error distribution of 0.29%).
• A power efficiency exceeding 8-trillion operations per second per Watt was demonstrated.

Conclusions:

• The developed HfSe2 memristor CBA offers a promising solution for energy-efficient and high-speed neural network hardware.
• The device enables hardware convolution image processing with programmable kernels, crucial for advanced neural network applications.
• This work overcomes key fabrication challenges, paving the way for practical 2D material-based neuromorphic computing.