Decoupling Dynamic Memory and Instantaneous Non-Linearity: A Hybrid Recurrent Architecture for Power Amplifier Modeling

The performance of next-generation wireless communication systems is critically limited by non-linear distortions in radio frequency power amplifiers (PAs). Digital predistortion (DPD) based on neural networks is a key enabling technology, yet existing models often struggle with a trade-off between accuracy and complexity. Here, we introduce the eXtended Recurrent Re-normalization Unit (X-RRU), a novel hybrid architecture that enhances a state-of-the-art recurrent neural network with a lightweight, direct feature bypass path. This path extracts physically informed, instantaneous features from the raw input signal, such as amplitude powers, and fuses them with the deep memory representation learned by the recurrent core. By decoupling the modeling of dynamic memory from instantaneous non-linearities, the X-RRU achieves superior performance with minimal added parameters. On a public 200 MHz DPD dataset, our model sets a new state-of-the-art with Normalized Mean Squared Error of −46.23dB and an Adjacent Channel Power Ratio of −52.25dBc. This work demonstrates that augmenting deep learning models with low-overhead physical features is a powerful and efficient strategy for mitigating non-linear distortions.