Surrogate Modeling With Complex-Valued Neural Nets for Signal Integrity Applications
Surrogate Modeling With Complex-Valued Neural Nets for Signal Integrity Applications
复制标题
DOI：
10.1109/tmtt.2023.3319835
复制
发表时间：
2024-01
影响因子：
4.3
通讯作者：
O. Akinwande;Serhat Erdogan;Rahul Kumar;Madhavan Swaminathan
中科院分区：
文献类型：
--
作者：
O. Akinwande;Serhat Erdogan;Rahul Kumar;Madhavan Swaminathan
神经网络(NN)在为许多信号完整性(SI)应用创建代理模型方面具有相当大的吸引力。基于神经网络的代理模型具有缩短设计周期时间和为设计者提供能够有效分析SI任务性能的快速原型的优点。因此，本文提出了一种新的端到端学习方法，用于使用复值神经网络进行代理建模，包含了更高的功能和更好的表示。该方法引入了深度复杂密集网络($\mathbb{C}$DNet)，该网络由复杂密集块构建以支持使用复值权重的复杂操作，并引入物理上一致的层来实施被动性和因果关系约束。提出了一种稳健的逆多目标优化方法，以最小化建模误差和优化设计空间参数。结果表明，在对相对较少的数据量进行正向学习和反向学习时，我们的模型优于最先进的深度代理模型。通过两个SI设计应用，验证了该方法的有效性，该模型用于预测宽带$S参数，并在给定期望目标规格的情况下获得最优设计空间参数。
Neural networks (NNs) are quite attractive in creating surrogate models for many signal integrity (SI) applications. NN-based surrogate models offer the benefits of reducing the design cycle time and providing the designer with a quick prototype that can efficiently analyze the performance of the SI task. This article, therefore, proposes a new end-to-end learning approach for surrogate modeling using complex-valued NNs, incorporating higher functionality and better representation. This approach introduces a deep complex dense network ( $\mathbb {C}$ DNet), which is built with complex dense blocks to support complex operations using complex-valued weights, and a physically consistent layer to enforce passivity and causality constraints. We also present a robust inverse multiobjective optimization method to minimize the modeling error and optimize the design space parameters. The results show that our model outperforms state-of-the-art deep surrogate models when tasked with forward and inverse learning for a relatively small amount of data. The effectiveness of the proposed approach is demonstrated through two SI design applications, where the model is used to predict broadband $S$ -parameters and obtain optimal design space parameters given the desired target specifications.