Hardware-Efficient Neural Network Digital Predistortion for Terahertz Power Amplifiers Using DeepShift and Pruning

With the increasing demand for high data rates in wireless communication systems, broadband signals with high peak-to-average power ratio (PAPR) are being utilized, and higher frequency bands such as terahertz waves are being adopted. These factors have led to significant nonlinear distortion in radio frequency (RF) power amplifiers (PAs). Particularly in high-frequency bands, PA nonlinearity becomes more pronounced, making it difficult to maintain signal quality. In this research, we propose a novel hardware-efficient neural network-based digital predistortion (DPD) technique that combines DeepShift and pruning. Our comprehensive evaluation using terahertz wave (256.4 GHz) PAs demonstrates that the proposed method, particularly when applying pruning before DeepShift (Strategy 1), maintains equivalent linearization performance with an EVM of 13.76% compared to 14.21% for conventional 32-bit fixed-point implementation, while reducing circuit area by 99.2% and energy consumption by 99.3% in the 45nm CMOS process. This remarkable hardware efficiency improvement without performance compromise is achieved through the synergistic combination of bit-shift operations replacing multiplications and strategic pruning of redundant connections. The proposed technique is particularly effective for PAs with complex nonlinearity in the terahertz band, providing an efficient DPD solution for high-performance linearization in resource-constrained environments.