Topology-Evolving Neural Network for Digital Predistortion of RF Power Amplifiers

DOI number:10.1109/TMTT.2026.3710476

Affiliation of Author(s):中国科学技术大学信息学院，东方理工大学

Journal:IEEE Transactions on Micriwave Theory and Techniques (Early Access)

Funded by:NSFC 62371436 等

Key Words:Digital predistortion (DPD), evolutionary algorithm, neural networks (NNs), power amplifier (PA).

Abstract:In scenarios where both low computational complexity and high linearization performance are strictly required, this article proposes a novel topology-evolving neural network(TENN) for digital predistortion (DPD) of radio frequency (RF) power amplifiers (PAs). Unlike conventional neural network (NN) models with both fixed architectures and predefined connections, the proposed TENN employs an evolutionary algorithm to automatically optimize the network topology. Starting from a minimal structure, the model progressively evolves into an efficient architecture with a specifically designed training strat egy. The proposed TENN incorporates two key mechanisms. First, mutation mechanisms are utilized to explore the optimal backbone structure of the model. Second, a tailored training strategy is introduced to progressively establish internal con nections according to their importance, enabling the TENN to adaptively refine its representational capability while main taining low complexity. Experimental results demonstrate that the proposed TENN achieves superior linearization performance compared with several state-of-the-art NN-based and classic linear-in-parameter DPD models, while requiring much lower computational complexity.

First Author:Junsen Wang (王俊森)

Co-author:Chengye Jiang,Renlong Han,Hao Chang,Qianqian Zhang

Indexed by:Journal paper

Correspondence Author:Kang Zhou,Falin Liu

Document Code:10.1109/TMTT.2026.3710476

Discipline:Engineering

Document Type:J

Volume:Online

Issue:Online

Page Number:1-14

Translation or Not:no

Date of Publication:2026-06-28