Multidimensional Load-Pull Extrapolation of Unknown Devices Using Generative Adversarial Networks

Nonlinear amplifier designers benefit from fast, accurate performance assessment and optimization techniques for computer-aided design. We demonstrate the synthesis of missing measurement dataset portions, even when extrapolating data beyond the boundary of the partial dataset’s domain. A novel method is presented to optimize device load impedance in conjunction with a third parameter. We demonstrate that this method can optimize settings for generic devices where an a priori estimate of the data range is unknown. A modified neural network architecture is used for improved data quality, and Bayesian optimization is applied to infer proper neural network normalization parameters for an unknown device. After applying the learned normalization parameters, the extrapolated dataset realizes a predicted optimal reflection coefficient per input power within a vector distance of about 0.1 on the Smith chart, with a mean output power error of less than 0.5 dB. This was accomplished using only 32 measurements—about 0.1% of the full set of 24192 measurements. If an a priori estimate of the normalization parameters is available, similar performance can be achieved in as few as 16 measurements. This is expected to be a useful tool for reducing design time and improving optimization accuracy in the computer-aided design of transistor amplifiers.