This article proposes a novel parallel gradient-based electromagnetic (EM) optimization approach to microwave components using adjoint-sensitivity-based neuro-transfer function (neuro-TF) surrogate. In the proposed technique, the surrogate model is trained using not only the input–output behavior but also the adjoint sensitivity information generated from the EM simulation simultaneously. By exploiting adjoint EM sensitivity for surrogate modeling, the proposed technique can obtain accurate surrogate models with larger valid range using the same amount of fine model evaluations compared with the existing gradient-based surrogate optimization without adjoint sensitivity. Furthermore, because the surrogate model is developed using adjoint EM sensitivity, the gradients calculated using the developed surrogate model in the proposed technique are much more accurate. The accurate gradients lead to further speedup of the surrogate optimization and improved quality of surrogate optimal solution in each surrogate optimization iteration. Since the surrogate model is valid in a large neighborhood and the gradients are sufficiently accurate, the proposed technique can achieve the optimal EM solution faster than the existing gradient-based surrogate optimization without adjoint sensitivity. Three examples of EM optimizations of microwave components are used to demonstrate the proposed technique.

中文翻译：

---

使用基于伴随灵敏度的神经传递函数代理对微波组件进行基于并行梯度的 EM 优化

本文提出了一种新的基于并行梯度的电磁 (EM) 优化方法，该方法使用基于伴随灵敏度的神经传递函数 (neuro-TF) 代理来优化微波组件。在所提出的技术中，代理模型不仅使用输入-输出行为进行训练，而且同时使用从 EM 模拟生成的伴随灵敏度信息。通过利用伴随 EM 敏感性进行代理建模，与现有的基于梯度的代理优化相比，所提出的技术可以使用相同数量的精细模型评估获得具有更大有效范围的准确代理模型。此外，由于代理模型是使用伴随 EM 灵敏度开发的，在所提出的技术中使用开发的替代模型计算的梯度要准确得多。准确的梯度导致代理优化的进一步加速和每个代理优化迭代中代理最优解的改进的质量。由于代理模型在大邻域中有效并且梯度足够准确，因此所提出的技术可以比现有的基于梯度的代理优化更快地获得最佳 EM 解决方案，而没有伴随敏感性。微波组件的 EM 优化的三个示例用于演示所提出的技术。由于代理模型在大邻域中有效并且梯度足够准确，因此所提出的技术可以比现有的基于梯度的代理优化更快地获得最佳 EM 解决方案，而没有伴随敏感性。微波组件的 EM 优化的三个示例用于演示所提出的技术。由于代理模型在大邻域中有效并且梯度足够准确，因此所提出的技术可以比现有的基于梯度的代理优化更快地获得最佳 EM 解决方案，而没有伴随敏感性。微波组件的 EM 优化的三个示例用于演示所提出的技术。