To enhance the efficiency and quality of communications, it is crucial to employ digital pre-distortion (DPD) technology for linearizing the power amplifiers (PAs) in terahertz/mmWave transceivers. Previous studies have shown that training DPD models within the direct learning architecture (DLA) framework yields superior results, as the Transformer-based PA behavioral model in this framework can directly compute the inverse function. Owing to resource-constrained deployment environments, DPD models trained via DLA typically rely on alternative lightweight models—such as long short-term memory (LSTM) networks. However, in scenarios where only DLA is applicable, lightweight DPD models often suffer from limited linearization performance. To mitigate this limitation, drawing inspiration from related work on iterative learning control (ILC), we propose a simple yet effective cross-architecture knowledge distillation method in the DLA framework (dubbed CAKDDLA). In this method, the lightweight DPD model is trained using two sources: the PA behavioral model and input-output knowledge distilled from a teacher model. To fully verify the proposed method’s effectiveness, extensive experiments conducted on a 1-GHz dataset show that our approach outperforms baseline methods in terms of error vector magnitude (EVM) and adjacent channel leakage ratio (ACLR), while also raising the upper bound of model performance.

To design reliable and efficient terahertz/mmWave transceivers, accurate power amplifier (PA) behavioral modeling is essential. In terahertz/millimeter wireless communication, the bandwidth will be 1 GHz or even more than 1 GHz, where PAs exhibit strong non-linearity and strong memory effects. This necessitates a powerful model capable of capturing long-term signal dependencies while managing strong non-linearity. To this end, we propose multi-scale augmented Transformer (MSAformer), which combines the ability of long short-term memory (LSTM) to capture complex sequence patterns with the self-attention mechanism’s dynamic attention adjustment, allowing it to effectively capture the intricate relationships between PA input and output signals. To fully validate the methods’ effectiveness, we collected and analyzed signals from the physical platform of the D-band system and set the bandwidth from 1 GHz to 4 GHz. Extensive behavioral modeling experiments on the collected datasets demonstrate that our method outperforms the existing methods in terms of normalized mean square error (NMSE). Further application in DPD scenarios proves that our method can effectively help improve the linearization performance of lightweight DPD models in terms of error vector magnitude (EVM) and adjacent channel leakage ratio (ACLR).

To ensure high-quality communication, it’s of great value to use digital pre-distortion (DPD) to linearize the core component power amplifier (PA) of terahertz/mmWave transceiver. In this work, we propose transfer learning with Transformer and LSTM for DPD of terahertz/mmWave transceiver, which uses Transformer based PA behavioral model to train effective and lightweight LSTM model for DPD. To collect and analyze the signal data of terahertz/mmWave transceiver, we set up a physical platform for the D-band system. Through experiments, we show that the proposed method is capable of significantly reducing in-band and out-band distortion while avoiding the excessive complexity of DPD models.

To meet the demand for high-speed and high-quality communication in next 6G satellite communication, it is very necessary and urgent to study the behavioral modeling of 6G satellite communication power amplifiers (PAs). In satellite communication, PAs face the situation of high dynamic and wide bandwidth and exhibit strong nonlinearity and strong memory effects. In this case, we need to study Transformer architectures that can better handle long sequence data and further explore the inherent characteristics of the PA signal data. In this article, we propose a behavioral modeling method of PAs named augmented real-valued time-delay transformer (ARVTDform). ARVTDform is an augmented transformer-based method, which can capture long-range dependencies between the PA signal data and has powerful nonlinear modeling capabilities. To simulate the working status of the satellite PAs, we set up two physical platforms and collect and analyze twelve datasets. To the best of our knowledge, this is the first time real satellite data has been used for behavioral modeling. Extensive experiments on the collected datasets further demonstrate that our transformer-based method is more suitable for handing PAs with strong nonlinearity and strong memory effects in terms of normalized mean square error (NMSE). Finally, we discuss the major challenge and list the potential future work that may contribute to the sustained development of high-performance transceivers.