Deep Learning-Assisted Design of Doherty Power Amplifiers Using Discretized Topology

This letter proposes an automated Doherty power amplifier (DPA) design methodology integrating theoretical S-parameter derivation with multiport artificial intelligence (AI) modeling. Specifically, this approach formulates DPA load-modulation requirements into explicit S-parameter targets, establishes a convolutional neural network (CNN) surrogate model for high-precision mapping, and optimizes the discretized topology to match these constraints. This strategy extends the scope of AI-assisted design, generalizing automated synthesis capabilities to complex load-modulated architectures. Afabricated 2.5–2.7 GHz DPA validates this methodology, demonstrating a saturation drain efficiency (DE) of 74%–81% at 43 dBm output power and a DE of 62%–71% at 6-dB output back-off (OBO). With digital predistortion (DPD), the design achieves an adjacent channel leakage ratio (ACLR) of −48.3 dBc under a 40-MHz 256-QAM signal with 6.3-dB peak-to-average power ratio (PAPR).