In-Memory Computing using Memristor Arrays with Ultrathin 2D PdSeO_x /PdSe₂ Heterostructure

Area of Science:

• Materials Science
• Computer Engineering
• Nanotechnology

Background:

• Classical von Neumann computing faces speed and energy limitations.
• Memristor arrays offer in-memory computing potential but suffer from ion transport variability.
• This variability compromises learning accuracy in neural network hardware accelerators.

Purpose of the Study:

• To address the variability issues in memristor-based in-memory computing.
• To develop a low-voltage memristor array with enhanced uniformity and reliability.
• To demonstrate the potential of this new memristor for neural network applications.

Main Methods:

• Fabrication of an ultrathin PdSeO_x /PdSe₂ heterostructure switching medium.
• Utilized controllable ultraviolet (UV)-ozone treatment for material realization.
• Investigated a distinct ion transport mechanism within the heterostructure.

Main Results:

• Achieved highly uniform memristor switching with low set (4.8%) and reset (-3.6%) voltage variability.
• Demonstrated confinement of conductive filament formation due to the unique ion transport mechanism.
• Implemented convolutional image processing with high recognition accuracy (≈93.4%) using crossbar kernels.

Conclusions:

• The developed PdSeO_x /PdSe₂ heterostructure memristor array overcomes conventional limitations.
• The device exhibits excellent uniformity and analog weight update characteristics for neural networks.
• This technology shows significant promise for efficient computing beyond the von Neumann architecture.