PulseRF: Physics Augmented ML Modeling and Synthesis for High-Frequency RFIC Design

The demand for highly efficient and compact passive networks is ubiquitous in modern radio frequency integrated circuits (RFICs). Traditional RF passive designs rely on a limited set of templates and require iterative manual efforts with compute-intensive simulations. This conventional approach restricts the design space and consumes significant time and resources. This work proposes to revolutionize the manual design process through a new framework, PulseRF. PulseRF efficiently automates the RF passive network design to achieve low-loss impedance transformation. It employs a physics-augmented UNet-based machine learning (ML) surrogate model to replace EM simulators. Experimental results demonstrate that the proposed physics-augmented ML model is capable of predicting the S-parameter matrix for a multi-layer 6-port passive network up to 300GHz. Bayesian optimization (BO) based design synthesis outperforms expert manual design with orders-of-magnitude speedup. Furthermore, PulseRF leads to the discovery of novel designs by efficiently exploring a broader design space that is incomprehensible to manual methods.