Scholarlink

<p>The design and optimization of epitaxial doping profiles for heterojuction bipolar transistorsrequire expert know-how and extensive trial-and-error using technology computer-aided design (TCAD) simulations. The vastness of the design exploration space along with time-intensive device simulation quickly renders conventional approaches infeasible. In this work, we propose a data-driven inverse design framework based on a surrogate Bayesian neural network model that not only predicts device performance with high accuracy, but also provides uncertainty estimates for each prediction. Leveraging these uncertainties, we implement an active learning scheme to iteratively refine the surrogate model by selecting the most informative new simulation points and thus ensuring maximal performance gain with minimal simulation times. Our results demonstrate that active learning leads to <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(&gt;45\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>&gt;</mo> <mn>45</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> improvement in network performance over random optimization approaches—paving a promising path to automated and efficient optimization. Finally, the efficacy of the approach is showcased by obtaining candidate epitaxial profiles through the framework that surpass expert-tuned state-of-the-art epitaxial designs for linearity and gain figure-of-merits.</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

Inverse design and optimization of heterojunction bipolar transistor epitaxy using active learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月10日

查看全文
  • A. N. M. Nafiul Islam,
  • Kai H. Kwok,
  • Cristian Cismaru

摘要

The design and optimization of epitaxial doping profiles for heterojuction bipolar transistorsrequire expert know-how and extensive trial-and-error using technology computer-aided design (TCAD) simulations. The vastness of the design exploration space along with time-intensive device simulation quickly renders conventional approaches infeasible. In this work, we propose a data-driven inverse design framework based on a surrogate Bayesian neural network model that not only predicts device performance with high accuracy, but also provides uncertainty estimates for each prediction. Leveraging these uncertainties, we implement an active learning scheme to iteratively refine the surrogate model by selecting the most informative new simulation points and thus ensuring maximal performance gain with minimal simulation times. Our results demonstrate that active learning leads to \(>45\%\) > 45 % improvement in network performance over random optimization approaches—paving a promising path to automated and efficient optimization. Finally, the efficacy of the approach is showcased by obtaining candidate epitaxial profiles through the framework that surpass expert-tuned state-of-the-art epitaxial designs for linearity and gain figure-of-merits.