Scholarlink

<p>Gallium nitride high-electron-mobility transistors (GaN HEMTs) play a crucial role in modern power electronics and high-frequency communications due to their superior electrical properties, including excellent breakdown voltage, electron mobility, and power density characteristics. Accurate modeling is essential for device optimization and performance prediction. While hybrid models combining physical and neural network approaches have gained attention due to their interpretability and extrapolation capabilities, they often require complex computations and numerous physical parameters. This study presents a hybrid modeling approach that effectively combines empirical models with neural networks, achieving both high accuracy and interpretability. This solution addresses key limitations: empirical models work well in linear regions but fail in saturation regions, while neural networks successfully correct these nonlinear effects. The proposed model incorporates a novel smoothness-based optimization that determines 14 neurons as the optimal configuration, preventing overfitting while maintaining physical interpretability through parameter consistency with conventional models. This work provides a practical framework for efficient GaN HEMTs characterization and application, offering significant improvements over existing methods in terms of both computational efficiency and modeling precision.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid neural network and empirical-based model for accurate I-V prediction of GaN HEMTs

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Zichen Zhang,
  • Xiaoying Xi,
  • Wei Huang,
  • Shuchao Zhang,
  • Guocheng Zhou

摘要

Gallium nitride high-electron-mobility transistors (GaN HEMTs) play a crucial role in modern power electronics and high-frequency communications due to their superior electrical properties, including excellent breakdown voltage, electron mobility, and power density characteristics. Accurate modeling is essential for device optimization and performance prediction. While hybrid models combining physical and neural network approaches have gained attention due to their interpretability and extrapolation capabilities, they often require complex computations and numerous physical parameters. This study presents a hybrid modeling approach that effectively combines empirical models with neural networks, achieving both high accuracy and interpretability. This solution addresses key limitations: empirical models work well in linear regions but fail in saturation regions, while neural networks successfully correct these nonlinear effects. The proposed model incorporates a novel smoothness-based optimization that determines 14 neurons as the optimal configuration, preventing overfitting while maintaining physical interpretability through parameter consistency with conventional models. This work provides a practical framework for efficient GaN HEMTs characterization and application, offering significant improvements over existing methods in terms of both computational efficiency and modeling precision.