<p>In this work, for the first time, an ensembled machine learning-based Hybrid Stacking approach is presented for small-signal behavioral modeling of high electron mobility transistors (HEMT). The device under test (DUT) is AlGaN/InGaN/GaN HEMT on a silicon carbide (SiC) substrate characterized at frequencies up to 50 GHz under room temperature. The stacking model was developed and trained on technology computer-aided design (TCAD)-generated data using four input parameters. It focuses on representing the device’s input–output behavior without delving deeply into the underlying physics. It can handle complex, nonlinear relationships and provide insights into device performance across varying conditions. The model’s predicted and simulated S-parameters show excellent agreement across the entire frequency range. The model demonstrated exceptional accuracy in both interpolation and extrapolation tests, achieving a mean absolute error (MAE) of 3.55E<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40042_2025_1416_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$-$</EquationSource> <EquationSource Format="MATHML"><math> <mo>-</mo> </math></EquationSource> </InlineEquation>03, mean squared error (MSE) of 5.20E<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40042_2025_1416_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$-$</EquationSource> <EquationSource Format="MATHML"><math> <mo>-</mo> </math></EquationSource> </InlineEquation>5, and root mean square error (RMSE) of 5.298E<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40042_2025_1416_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$-$</EquationSource> <EquationSource Format="MATHML"><math> <mo>-</mo> </math></EquationSource> </InlineEquation>03. The R-squared and explained variance scores were approximately 0.99 and 0.998, respectively. By precisely capturing the dependability of S-parameters on bias points and operating conditions, the proposed methodology highlights its potential to reduce barriers to adopting machine learning techniques in semiconductor research. This approach enhances the understanding of GaN HEMT performance and encourages the exploration of advanced ML models for broader applications in device analysis and optimization.</p>

Data-driven small-signal modeling of AlGaN/InGaN/GaN high electron mobility transistor using multi-layered ensemble learning

期刊论文
发表日期：

2025年6月17日

查看全文

Neda Ahmad,
Sonam Rewari,
Vandana Nath

摘要

In this work, for the first time, an ensembled machine learning-based Hybrid Stacking approach is presented for small-signal behavioral modeling of high electron mobility transistors (HEMT). The device under test (DUT) is AlGaN/InGaN/GaN HEMT on a silicon carbide (SiC) substrate characterized at frequencies up to 50 GHz under room temperature. The stacking model was developed and trained on technology computer-aided design (TCAD)-generated data using four input parameters. It focuses on representing the device’s input–output behavior without delving deeply into the underlying physics. It can handle complex, nonlinear relationships and provide insights into device performance across varying conditions. The model’s predicted and simulated S-parameters show excellent agreement across the entire frequency range. The model demonstrated exceptional accuracy in both interpolation and extrapolation tests, achieving a mean absolute error (MAE) of 3.55E $-$ $-$ 03, mean squared error (MSE) of 5.20E $-$ $-$ 5, and root mean square error (RMSE) of 5.298E $-$ $-$ 03. The R-squared and explained variance scores were approximately 0.99 and 0.998, respectively. By precisely capturing the dependability of S-parameters on bias points and operating conditions, the proposed methodology highlights its potential to reduce barriers to adopting machine learning techniques in semiconductor research. This approach enhances the understanding of GaN HEMT performance and encourages the exploration of advanced ML models for broader applications in device analysis and optimization.