Multimodal ion-gated transistor based on 2D superionic conductor for in-memory computing in deep learning

Area of Science:

• Materials Science
• Device Physics
• Artificial Intelligence Hardware

Background:

• Neuromorphic computing aims to mimic the brain's efficiency by integrating computational functions.
• Simultaneously achieving linear multiply-accumulate and nonlinear activation in a single device presents significant material and engineering challenges.
• Existing approaches often require separate components, increasing latency and power consumption.

Purpose of the Study:

• To design and fabricate a single device capable of performing both multiply-accumulate and nonlinear activation operations for neuromorphic computing.
• To overcome the conflicting material requirements for these two distinct functions.
• To demonstrate the device's potential for efficient AI hardware applications.

Main Methods:

• Fabrication of a multimodal ion-gate transistor utilizing a 2D CdPS₃-Li dielectric layer and MoS₂ channel.
• Exploitation of the layered CdPS₃-Li structure for anisotropic ion transport and Li⁺ storage.
• Leveraging Cd vacancies in CdPS₃-Li to trap photo-generated holes from MoS₂ for nonlinear activation.

Main Results:

• The fabricated transistor exhibited high-linearity and non-volatile resistance states under electrical pulses due to strong ion-electron coupling.
• The device demonstrated rich nonlinear behavior under light pulses, attributed to Cd vacancies interacting with photo-generated holes.
• The CdPS₃-Li transistor arrays achieved high accuracy in handwritten digit classification tasks.

Conclusions:

• A single transistor device capable of simultaneously performing multiply-accumulate and nonlinear activation operations has been successfully demonstrated.
• The unique properties of the 2D CdPS₃-Li/MoS₂ heterostructure enable the integration of conflicting functionalities.
• This multimodal transistor offers a promising hardware solution for low-latency, high-efficiency neuromorphic computing systems.