The shift toward 5G networks and future generations imposes increasingly stringent demands on power amplifiers, especially in terms of efficiency over wide back-off ranges. The classic Doherty amplifier remains a reference architecture, but its traditional design with fixed combiners limits the ability to maintain optimal performance when operating with signals of high peak-to-average power ratio. A novel approach, recently developed, combines pixelated combining networks with deep learning-based surrogate models to overcome these barriers. Instead of relying on costly electromagnetic simulations, a deep convolutional neural network is trained to quickly and accurately predict the S-parameters of such passive networks. This surrogate is integrated into a genetic algorithm optimization framework, enabling the synthesis of complex Doherty combiners that extend the efficient back-off range using fully symmetric devices. Prototypes fabricated with GaN HEMT transistors demonstrate a maximum drain efficiency above 74% and an output power exceeding 44.1 dBm at 2.75 GHz, maintaining over 52% efficiency at 9 dB back-off. When subjected to a 20 MHz 5G NR signal with a PAPR of 9 dB and applying digital predistortion, an average power-added efficiency above 51% and an ACLR better than -60.8 dBc are achieved.
The integration of artificial intelligence into radio frequency hardware design not only accelerates the process but also opens the door to optimizations that were previously unfeasible due to their computational cost. For these methodologies to reach the industry with guarantees, it is essential to have technology partners that offer artificial intelligence solutions for businesses and custom applications that enable the implementation of these complex workflows. Q2BSTUDIO, as a software and technology development company, provides exactly that ecosystem: from custom software for creating surrogate models to AWS and Azure cloud services that scale the necessary massive simulations. Its capabilities in business intelligence services, such as Power BI, facilitate the visualization of optimization results, while cybersecurity protects the intellectual property of the designs. Additionally, the development of custom AI agents allows for automating the exploration of combiner topologies, drastically reducing design times. This convergence between cutting-edge research and robust enterprise solutions is what makes concepts like the black-box Doherty amplifier with deep learning and pixelated combiner a commercial reality, driving the next generation of communications infrastructure.
