MOSFET Physics-Based Compact Model Mass-Produced: An Artificial Neural Network Approach.

Area of Science:

• Semiconductor device physics
• Computational physics
• Artificial intelligence in engineering

Background:

• Scaling down nanoscale semiconductor devices presents challenges in obtaining analytical surface potential solutions from complex physics equations.
• Accurate surface potential solutions are fundamental for MOSFET compact models.
• Traditional 1D Poisson equation solutions are insufficient for modern scaling devices.

Purpose of the Study:

• To propose a general framework for automatically deriving analytical surface potential solutions in MOSFETs.
• To leverage deep neural networks for handling complex physics equations and generating closed-form solutions.
• To improve the accuracy and applicability of MOSFET compact models.

Main Methods:

• Developed a framework utilizing deep neural networks, specifically a physical-relation-neural-network (PRNN).
• PRNN learned from a general-purpose numerical simulator to handle complex mathematical physics equations.
• Generated accurate closed-form mapping between device parameters and surface potential, reflecting 2D Poisson equation solutions.

Main Results:

• Successfully inferred the analytic surface potential of MOSFETs with high accuracy.
• Applied the derived potential function to build 130 nm MOSFET compact models.
• Demonstrated accurate prediction of device performances in circuit simulations.

Conclusions:

• The proposed deep learning framework efficiently derives analytical surface potential solutions for MOSFETs.
• The method surpasses traditional 1D solutions by reflecting 2D Poisson equation characteristics, better illustrating scaling device physics.
• This framework shows significant potential for device optimization and advanced circuit design.