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

Sammanfattning

AI-driven inverse design of RF circuits offers a promising approach to achieving ahigh degree of design automation. However, training an underlying neural networksurrogate typically requires a massive number of training samples. In addition, thesesurrogates often suffer from limited generalizability. For instance, redefining thedesign 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 developa fully open-source, multi-fidelity framework where a neural network surrogate ispre-trained on a large-scale dataset derived from fast analytical approximations,and subsequently refined via transfer learning using a sparse set of high-fidelity EMsimulations. This approach is validated by designing low-pass filters on a two-layerPCB with an 8×8 grid. The framework achieves target predictive accuracy whilereducing the required high-fidelity EM training samples from 10k to 200, representing a 50-fold reduction in training data requirement. Consequently, the total datageneration time drops from 83.3 hours to 5.2 hours, yielding a 16-fold computational speedup. This methodology provides a practical and data-efficient strategyto improve the overall efficiency of automated RF hardware design.