A Physics-Informed Automatic Neural Network Generation Framework for Emerging Device Modeling.

Area of Science:

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

Background:

• Traditional equation-based semiconductor modeling faces accuracy and time constraints.
• Neural network (NN)-based models offer potential but suffer from unphysical behaviors and complex structure optimization.
• Existing NN models lack smoothness and monotonicity, limiting practical application.

Purpose of the Study:

• To propose an Automatic Physical-Informed Neural Network (AutoPINN) generation framework.
• To resolve unphysical behaviors in NN-based compact models.
• To automate the determination of optimal NN structures for semiconductor device simulation.

Main Methods:

• Developed a framework combining Physics-Informed Neural Networks (PINNs) and a two-step Automatic Neural Network (AutoNN).
• PINNs incorporate physical information to ensure model validity.
• AutoNN automatically optimizes NN architecture without expert intervention.

Main Results:

• The AutoPINN framework achieved less than 0.05% error on gate-all-around transistor devices.
• Demonstrated preservation of smoothness in high-order derivatives and monotonicity.
• Validated promising generalization capabilities through test error and loss landscape analysis.

Conclusions:

• AutoPINN effectively resolves unphysical behaviors and automates NN structure selection.
• The framework accelerates the development and simulation of emerging semiconductor devices.
• This approach offers a robust and efficient alternative to traditional modeling methods.