The conventional way of mitigating distortion and correcting for nonlinearities in power amplifiers is through digital pre-distortion (DPD), where the nonlinearity of the device is captured so that an inverse model can be applied in the digital baseband. The most common DPD architectures rely on nonlinear behavioral models based on the Volterra series, such as the memory polynomial model (MPM) or generalized memory polynomial model (GMPM). These linearization methods have adequate performance but due to the wide bandwidth introduced in 5G Millimeter-Wave (up to 2 GHz) and requirements for power efficiency (which results in high nonlinearity), the computational complexity is extreme and needs a huge number of coefficients in the mathematical models. On the other hand, the hardware complexity of the active phased array (APA) transmitters in the 5G system introduces additional limitations for conventional measurement approaches and requires over-the-air (OTA) measurement.

Artificial neural networks (ANN) are well known to be able to learn any arbitrary nonlinear function according to the universal approximation theorem. When comparing the MPM approach with the ANN approach, the MPM has inherent local approximating properties in contrast to the global approximation capability of ANNs, when modeling strongly nonlinear systems. In addition, when compared to classical models, the ANN may adapt better to extrapolating beyond the zone exploited for parameter extraction.

This research mainly deals with defining specific design challenges for Millimeter-Wave APA architectures and proposing solutions and algorithms for linearity enhancement of APAs based on ANN together with advanced OTA measurement techniques for verification of the proposed methods.