Multidimensional Load-Pull Extrapolation using Generative Adversarial Networks
Jonathan E. Swindell, Adam C. Goad, Austin Egbert, Charles Baylis, Robert J. Marks · 2025
Design optimization of high-power nonlinear power amplifiers requires much back-and-forth between power sweep, bias sweep, and load-pull simulations. Such manual optimizations are very slow given the large number of parameters and goals. The automation of this process holds promise for reducing time to successful designs in computer-aided design operations, especially in microwave power-amplifier design. This work proposes multidimensional image completion techniques using Generative Adversarial Networks (GAN) as a method for reducing the number of simulation queries required to characterize a power amplifier over multiple design parameters simultaneously, such as input power and load impedance. Specifically, as few as 16 simulation queries are able to accurately predict the optimal reflection coefficient over input power to within a vector error of about 0.1, leading to significant time savings over other characterization methods.