Survival of the Fastest: Rapid 4-Port Passive RFIC Optimization via Surrogate-Assisted Genetic Algorithms

Abstract

The electrical behavior of Radio Frequency Integrated Circuits (RFICs) is strongly influenced by the layout, necessitating the use of full-wave electromagnetic simulation throughout the design process, which is both computationally expensive and prohibitively time-consuming. To address this issue, this thesis presents a fully open-source pipeline for surrogate-assisted inverse design of pixel grid RFIC structures, capable of producing designs with a maximum of 4 ports in under a minute. The workflow consists of three stages: dataset generation, surrogate model training, and finally design target optimization and generation. Utilizing the IHP SG13G2 130nm open PDK, the OpenEMS FDTD full-wave EM solver, and a llama-based parser and design interface, the entire workflow is open to the research community, eliminating the reliance on research-prohibitive proprietary EDA tools. In the dataset generation, we generate a dataset of 40K randomized 25 x 25 binary pixel-grid layouts with 1 port per bordering edge. The layouts are simulated across a range of frequencies to capture the S-parameter matrices for training a surrogate model. The surrogate, a 12-layer convolutional neural network, is then trained on the upper triangle of the S-parameter matrix to ensure reciprocity and passivity for the networks and reduce the amount of data being learned. To boost model performance, we augment the dataset with D4 symmetrical permutations, and use active learning to get real simulation data in the most prominent regions of the design space. A memetic genetic algorithm leverages the surrogate to guide a multi-stage evolutionary search to find designs that meet the user specifications in terms of S-parameters across different frequencies. To aid the user in inputting many complex-valued targets as design objectives, we use an open source LLM to parse the user's natural language input. Finally, to improve the design space search, we add a term to the GA cost function to reward consensus across models. This modular, open-source architecture serves as a baseline for automated RFIC design.