Physics Informed Machine Learning based Digital Predistortion of RF Power Amplifiers

Considering the environmental impact of radio access technologies alongside traditional performance metrics that have driven the evolution of 4G and 5G systems, it’s essential to recognize that wireless infrastructure consumes a significant portion of energy used by communication service providers. Several techniques, including resource optimization at the network level, traffic-driven power supply modulations, and various energy saving modes in the radio unit (RU) have shown to play a crucial role in reducing the energy consumption of a radio access network for different traffic loads. However, when the transmitter is pushed to operate at near maximum power ratings to enhance higher wireless channel capacity, machine learning (ML) based digital pre-distortion (DPD) techniques in conjunction with high-efficiency power amplifiers (PAs) become the cornerstone to the successful deployment of energy-efficient basestation RUs. In this presentation, we will explore the underlying technology, namely, adaptive DPD, and illustrate how ML techniques can be effectively incorporated into the design methodology of PA linearization. We will introduce the problem by analyzing the drain inductance resonance behavior of the PA, demonstrate the performance impact due to this undesired behavior, and share our findings. Particularly, our proposed solution features a physically informed neural network that can be efficiently mapped to a digital system architecture to address the new challenge in wideband basestation RUs. We hope the study's findings will illuminate the modeling capabilities of the state-of-the-art DPD research and provide insights into future directions from an industrial perspective.