Characteristics Prediction and Optimization of GaN CAVET Using a Novel Physics-Guided Machine Learning Method.

Area of Science:

• Semiconductor device physics
• Machine learning applications
• Materials science

Background:

• Gallium Nitride (GaN) current aperture vertical field effect transistors (CAVETs) are crucial for high-power electronics.
• Accurate modeling of CAVETs' I-V characteristics is essential for device optimization and circuit design.
• Existing modeling approaches may struggle with small datasets and ensuring physical interpretability.

Purpose of the Study:

• To develop a physics-guided machine learning (PGML) model for predicting GaN CAVET I-V characteristics.
• To leverage transfer learning and a novel neural network architecture for enhanced predictive accuracy.
• To ensure physically meaningful and robust predictions for practical applications.

Main Methods:

• Implementation of a physics-guided neural network incorporating transfer learning and a shortcut structure.
• Utilizing a shallow neural network with tanh basis functions and a hypernetwork for dynamic weight generation.
• Incorporating transconductance influence into the loss function for synchronous prediction of output and transfer characteristics.

Main Results:

• The PGML model achieved prediction errors within 5% on small sample datasets.
• The model demonstrated high accuracy, with R2 values exceeding 0.99.
• The approach successfully predicted both output and transfer characteristics synchronously.

Conclusions:

• The proposed PGML approach offers superior performance compared to conventional methods for GaN CAVET modeling.
• The model provides physically meaningful and robust predictions, suitable for device optimization and circuit-level simulations.
• This work demonstrates the potential of PGML in advancing semiconductor device modeling.