AI-Enabled Design of On-Chip Transformers: Transfer Learning for Inverse Synthesis Across RFIC Fabrication Processes

Abstract

Current techniques for Radio Frequency Integrated Chip (RFIC) design heavily rely on preexisting templates and trial-and-error methods, leading to significant process inefficiencies and hindering innovation in design. This project aims to address these challenges through AI-led tools which can automate aspects of the RFIC design cycle, specifically applied to on-chip transformers. By developing a machine learning (ML) model capable of synthesizing transformer geometric parameters from desired input-output behavior, the need for time-consuming electromagnetic (EM) simulations for each design iteration is greatly reduced. To achieve this inverse design workflow, a forward model must first be developed which maps on-chip transformer layout geometry to its resulting transmission and reflection characteristics at each port (S-parameters). However, generating a robust training dataset for each individual chip fabrication technology presents its own challenges. This project uses transfer learning techniques to maximize the accuracy of S-parameter predictions between distinct fabrication processes. This paper discusses experimental results for the forward training loop, as well as transfer learning optimizations between the 9HP and 22FDX fabrication technology datasets. Future research can build on this optimized forward model to finalize the inverse design workflow and facilitate its adoption in real-world industrial RFIC development.