Machine Learning for Evaluating the Impact of Manufacturing Process Variations in High-Speed Interconnects

This paper presents a machine learning based modeling methodology to analyze the impact of high-volume manufacturing process variations on electrical performance of high-speed interconnects, that overcomes the limitations of traditional approaches. The proposed methodology outperforms the response surface based modeling for high-speed interconnects and is capable of handling highly nonlinear relationships. Machine learning is demonstrated to be a promising approach to explore design spaces efficiently and accurately even when modeling data is limited due to expensive computational cost.

1. Introduction

Variation in a high-volume manufacturing (HVM) process is a major concern in the design of high-speed interconnects [1] –[3]. The ever-increasing demand for higher bandwidth and lower loss with shrinking design margins makes system performance even more susceptible to uncertainty. Without the means to evaluate signal behavior in a systematic manner as the physical and electrical characteristics of the system components vary, uncertainty may cause significant performance degradation and yield reduction or result in an overly designed system with increased design cycles and cost [3].