Publication: Performance analysis of ann, svr, and gbm models for s-parameter prediction of high-speed interconnects using different data configuration techniques
Loading...
Date
2024-07
Authors
Goh, Jun Hui
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
High-speed printed circuit boards (PCBs) enable rapid signal transmission but face challenges in maintaining signal and power integrity (SI/PI) at high frequencies, particularly issues like crosstalk. This research addresses this challenge by leveraging Machine Learning (ML) models to predict the S-parameters (crucial metrics used to characterize the electrical behavior) of complex PCB structures. This approach can enhance the efficiency and reliability of high-speed PCB designs. This study utilized a sizeable dataset from the SI/PI-Database, '5x5 Via-Array on a 10-Cavity PCB', to train various ML models such as Artificial Neural Network (ANN), Support Vector Regression (SVR), and Gradient Boosting Machine (GBM). The results demonstrated that the frequency-integrated many-to-one approach achieved the highest accuracy and efficiency, making it the optimal technique. Specifically, the ANN model with a [20 40 20] architecture delivered the most accurate S11 magnitude predictions (RMSE 0.0059, tolerance rate 97.08%, R-squared 0.9978) and minimal memory usage (0.0837 MB). Conversely, the GBM model with 1000 learning cycles and 0.2 learning rate achieved the fastest training time (90 seconds), suitable for time-constrained tasks, albeit with substantial memory usage. This study underscores the potential of optimized ANN models for efficient and reliable S-parameter prediction in high-speed PCBs, promoting broader AI-based circuit modeling and innovation.