Analog memristive synapse based on topotactic phase transition for high-performance neuromorphic computing and neural...

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Nonvolatile memories (NVMs) are crucial for power-efficient artificial intelligence (AI) hardware but face challenges with device variability.
• Neuromorphic computing aims to mimic the human brain's efficiency using NVMs.

Purpose of the Study:

• To demonstrate a novel Topotactic Phase Transition Random-Access Memory (TPT-RAM) with a dual-mode operation.
• To investigate the switching mechanism and reduce variability in NVMs for AI applications.
• To implement TPT-RAM for neural network pruning and improve image classification.

Main Methods:

• Fabrication and characterization of SrCoO-based TPT-RAM devices.
• Utilizing density functional theory (DFT) and kinetic Monte Carlo (KMC) simulations to study the switching mechanism.
• Implementing the dual-mode TPT-RAM in a neural network for image classification tasks.

Main Results:

• Demonstrated a SrCoO-based TPT-RAM with reproducible analog switching and reduced variability.
• Identified an orientation-dependent switching mechanism controlled by oxygen ion migration.
• Achieved ~84.2% reduction in redundant synapses and 99% image classification accuracy using the TPT-RAM.

Conclusions:

• TPT-RAM offers a promising solution for overcoming device variability in NVMs.
• The developed memristive synapse design provides a new pathway for bioplausible neuromorphic computing.
• This research advances the development of efficient AI hardware through novel materials and device architectures.