Scholarlink

Abstract <p>This paper presents a Deep Neural Network (DNN)-based modeling framework for accurately predicting the current–voltage (<i>I</i>–<i>V</i>) characteristics of AlInGaN/GaN High Electron Mobility Transistors (HEMTs). The proposed model leverages the nonlinear learning capability of DNNs to capture complex relationships between fabrication parameters. These include aluminum (<i>Al</i>) and indium (<i>In</i>) mole fractions, barrier thickness, and gate dimensions. The quaternary AlInGaN barrier layer improves polarization control and the formation of a two-dimensional electron gas (2DEG), which requires robust modeling for performance optimization. A three-layer DNN architecture was trained on experimental I–V datasets and validated against multiple published device structures. The model demonstrates excellent agreement with measured data, significantly outperforming conventional compact models in both accuracy and adaptability. This approach provides a cost-effective, scalable, and data-driven alternative for modeling advanced gallium nitride (GaN) HEMTs, offering improved convergence, reduced development time, and potential for broader application in GaN-based power and RF electronics. The model achieves an RMSE of 0.004 and an MAE of 0.011, outperforming previously reported machine learning and hybrid models.</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

Modeling of I–V Characteristics of AlInGaN/GaN HEMTs using Deep Neural Networks

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Kavita Thorat Upadhyay,
  • Prateek Nahar,
  • Arvind Upadhyay,
  • Abivyakt Bhati

摘要

Abstract

This paper presents a Deep Neural Network (DNN)-based modeling framework for accurately predicting the current–voltage (I–V) characteristics of AlInGaN/GaN High Electron Mobility Transistors (HEMTs). The proposed model leverages the nonlinear learning capability of DNNs to capture complex relationships between fabrication parameters. These include aluminum (Al) and indium (In) mole fractions, barrier thickness, and gate dimensions. The quaternary AlInGaN barrier layer improves polarization control and the formation of a two-dimensional electron gas (2DEG), which requires robust modeling for performance optimization. A three-layer DNN architecture was trained on experimental I–V datasets and validated against multiple published device structures. The model demonstrates excellent agreement with measured data, significantly outperforming conventional compact models in both accuracy and adaptability. This approach provides a cost-effective, scalable, and data-driven alternative for modeling advanced gallium nitride (GaN) HEMTs, offering improved convergence, reduced development time, and potential for broader application in GaN-based power and RF electronics. The model achieves an RMSE of 0.004 and an MAE of 0.011, outperforming previously reported machine learning and hybrid models.