In the context of process design kit (PDK) development, obtaining accurate broadband $S$ -parameters of interconnect structures is often challenged by limited measurement bandwidth and the high computational cost of wide-frequency electromagnetic (EM) simulations. To address these constraints, this article proposes a frequency-domain extrapolation method based on machine learning (ML). Unlike traditional approaches that rely on physical modeling or focus solely on local spectral patterns, the proposed multihead convolutional neural network-long short-term memory network transfer learning (MH-CNN-LSTM-TL) model establishes a nonlinear mapping between low- and high-frequency responses, with geometric parameters incorporated as auxiliary inputs to achieve accurate extrapolation of high-frequency $S$ -parameters. The model employs a multibranch architecture, enabling the parallel extrapolation of multiple $S$ -parameter frequency responses. To mitigate the high cost of EM simulation data acquisition, a transfer learning (TL) strategy from circuit simulation data to EM simulation data is introduced, which significantly improves generalization under small-sample conditions and reduces computational overhead. The effectiveness of the proposed approach is evaluated with three representative simulation cases: one on microstrip lines and two on complex through-silicon via (TSV) with redistribution layer (RDL) interconnect structures, in addition to validation with experimental data from differential microstrip lines. Results demonstrate superior prediction accuracy across a wide-frequency range, along with strong adaptability and generalization capability, highlighting its potential for practical engineering applications.
原文链接
分享
引用
