Augmented Real-Valued MLP-Based Digital Predistortion for Microwave PAs in ILA Framework

Abstract

This paper presents an augmented real-valued multi-layer perceptron (MLP)-based digital predistortion (DPD) technique for the linearization of wideband power amplifiers (PAs). Implemented within an indirect learning architecture (ILA), the proposed approach incorporates envelope-dependent augmented features together with Cartesian in-phase and quadrature (I/Q) components to effectively capture both odd- and even-order nonlinearities. The proposed methodology is validated through both behavioral simulations and hardware measurements, demonstrating robust performance under practical operating conditions. For 64-QAM signals, the proposed DPD improves constellation fidelity and reduces the error vector magnitude (EVM) from 3.7% to 2.8%. In high-order 256-QAM scenarios, it outperforms the conventional memory polynomial (MP) model, achieving a 2.4% absolute EVM reduction from 4.5% to 2.1%. Furthermore, under 20 MHz wideband excitation, the EVM is reduced from 4.8% to 2.6%. These results verify that the proposed augmented real-valued MLP-based DPD achieves effective and hardware-efficient linearization for wideband PAs supporting high-order modulation schemes.