A Modular Domain-Aware Neural Network Architecture for Power Amplifier Linearization

Next-generation wireless systems demand highly efficient power amplifiers (PAs), which introduce complex non-linearities and memory effects. Traditional digital pre-distortion (DPD) methods, such as the Generalized Memory Polynomial (GMP), struggle with these impairments. This paper proposes a Modular Domain-Aware Time-Delay Neural Network (MDA-TDNN), a hybrid physically-inspired architecture that combines established signal processing components into a unified TDNN-based structure. Rather than introducing fundamentally new building blocks, the contribution lies in the systematic integration and evaluation of these components, providing a structured characterization of their individual and joint effects on DPD performance. It includes a Finite Impulse Response (FIR) layer for memory modeling, a phase normalization layer for complex signal behavior, an envelope layer for amplitude non-linearities, and a residual path for stable training. MDA-TDNN was trained on an ideal dataset generated by an Iterative Learning Control scheme and experimentally validated on an RF testbed using a 75 MHz 5G NR signal. The results demonstrate that MDA-TDNN outperforms GMP by 7.95 dB in Adjacent Channel Leakage Ratio (ACLR), achieving performance within 1 dB of ideal DPD. An ablation study highlights the importance of phase normalization and FIR layers. These findings validate the effectiveness of domain-aware neural architectures for PA linearization.