PACOSYT: A Passive Component Synthesis Tool Based on Machine Learning and Tailored Modeling Strategies Towards Optimal RF and mm-Wave Circuit Designs

Abstract

In this paper, the application of regression-based supervised machine learning (ML) methods to the modeling of integrated inductors and transformers is examined. Different ML techniques are used and compared to improve accuracy. However, it is demonstrated that none of the ML techniques considered provided good results unless a smart modeling strategy, tailored to the specific design problem, is used. Taking advantage of these modeling strategies, high accuracy can be obtained when compared to full-wave electromagnetic (EM) simulations (less than 2% error) and experimental measurements (less than 5% error). The most accurate model, obtained by the appropriate combination of an ML technique and modeling strategy, has been integrated into a tool called PACOSYT. The tool uses optimization algorithms to allow the designer to obtain an inductor/transformer with optimal performances in just seconds while keeping the accuracy of EM simulations. Furthermore, the tool provides the passive component S parameter description file for seamless use in circuit simulations. The tool can be used standalone or integrated with design frameworks, like Cadence Virtuoso or AIDASoft, a framework for circuit optimization. To illustrate the different usages of the tool, several passive devices are synthesized, and hundreds of millimeter-wave power amplifiers are synthesized using AIDASoft together with PACOSYT. The tool has been developed using open-source Python frameworks and does not use any closed-source licenses. PACOSYT, which also allows other designers to create their models for different technologies, is made publicly available.

Keywords

• Design automation
• inductors
• integrated circuits
• machine learning
• millimeter-wave
• modeling