Empowering Optimal Design of RF Devices by Generative AI

Traditional generative design was based on parametrized physics (including domain geometry) on which optimization procedures operate. In order to alleviate the cost of 3D high-fidelity simulations model order reduction technologies first, and neural-based surrogates were proposed, enabling real-time queries. However, sometimes the addressed configurations make difficult an efficient parametrization (the number of parameters can grow exponentially) because of the so-called combinatorial exposition or curse of dimensionality. In those situations, both the construction of the surrogate associated to a convenable always challenging design of experiments, as well as the exploration of the multi-parametric solution cannot be envisaged. Recent technologies of generative AI, at the heart of LLM, open appealing perspectives for circumventing the just referred limitations. By using appropriate autoencoders (as the ones our group recently proposed) the amount of data can be drastically reduced, the descriptors discovered in an optimal way, to explain the observed physics to be modeled, while keeping excellent properties of generation and interpolation. These discovering-while-learning technologies enable extrapolation, and can proceed from a diversity of data, in particular fields, graphs, tabulated data, … where transformer with attention mechanisms embedded into a linear reduced latent space where optimization, interpolation and generation proceed. We introduce in this presentation a Rank-Reduction AutoEncoder (RRAE) framework as a generative surrogate modeling approach for the design of RF circuits. A few examples of 2D and 3D devices are presented as well as antennas. S-parameters and radiation patterns associated to optimized designs are described and demonstrate the efficiency of generative AI.