Data-Efficient AI-Driven RF Circuit Design via Multi-Fidelity Surrogates

Abstract

AI-driven inverse design of RF circuits offers a promising approach to achieving a high degree of design automation. However, training an underlying neural network surrogate typically requires a massive number of training samples. In addition, these surrogates often suffer from limited generalizability. For instance, redefining the design space often requires completely rebuilding the full-wave electromagnetic simulation dataset. Thus, data generation remains a major computational bottleneck. This thesis investigates how to mitigate this data-demanding challenge. We develop a fully open-source, multi-fidelity framework where a neural network surrogate is pre-trained on a large-scale dataset derived from fast analytical approximations, and subsequently refined via transfer learning using a sparse set of high-fidelity EM simulations. This approach is validated by designing low-pass filters on a two-layer PCB with an 8×8 grid. The framework achieves target predictive accuracy while reducing the required high-fidelity EM training samples from 10k to 200, representing a 50-fold reduction in training data requirement. Consequently, the total data generation time drops from 83.3 hours to 5.2 hours, yielding a 16-fold computational speedup. This methodology provides a practical and data-efficient strategy to improve the overall efficiency of automated RF hardware design.