A Physics-Informed 1D-CNN Surrogate Model Using Geometric Features for Mode Conversion Prediction in High-Density C-PHY Routing
Journal
30th IEEE Workshop on Signal and Power Integrity, SPI 2026
Start Page
1
End Page
4
ISBN
[9798331590758]
Date Issued
2026-06-14
Author(s)
Huang, Tse-Jui
Abstract
In high-speed mobile interfaces, such as MIPI CPHY, inter-lane skew matching is critical for maintaining signal integrity. To meet constraints within compact routing areas, designers often employ tightly coupled serpentine delay lines. However, the geometric asymmetries and bending discontinuities in these structures inevitably induce differential-to-common mode noise, which degrades eye margins and introduces electromagnetic interference (EMI) issues. Traditional full-wave electromagnetic (EM) simulations are computationally prohibitive for iterative optimization, while existing image-based deep learning methods suffer from physical size extracting limitations and lack physical interpretability. This paper proposes a computationally efficient, physics-informed surrogate model based on a 1Dimensional Convolutional Neural Network (1D-CNN). Instead of processing raw layout images, the proposed method extracts nine critical geometric features to explicitly quantify the electromagnetic coupling and impedance discontinuities along the path. Trained by $3,600 \text{HFSS}$-simulated datasets, the model achieves a Root Mean Square Error (RMSE) of approximately 2 dB with an inference speedup of several orders of magnitude compared to full-wave solvers. This lightweight and accurate surrogate model enables real-time performance prediction, enabling automated layout synthesis and reinforcement learning-based design optimization.
Event(s)
30th IEEE Workshop on Signal and Power Integrity, SPI 2026
Subjects
C-PHY
Convolutional Neural Network
Delay lines
Machine learning
Mode conversion
Routing
Surrogate model
Publisher
IEEE
Type
conference paper