A Low-Complexity Plug-and-Play Deep Learning Model for Generalizable Massive MIMO Precoding

TL;DR

This work addresses the practical challenge of efficient, robust downlink precoding in massive MIMO. It introduces PaPP, a plug-and-play backbone with a teacher–student, self-supervised, meta-learning framework that generalizes across deployment sites, transmit powers, and CSI quality, and can operate in both fully digital and hybrid beamforming modes. The approach yields substantial energy savings (over 21x) and maintains high spectral efficiency, even under CSI errors, with the option to fine-tune on local unlabeled data using a small sample set. Experimental results on realistic ray-tracing channels demonstrate that PaPP outperforms conventional baselines and prior learning-based methods, offering a practical, deployment-friendly path toward energy-efficient mMIMO precoding.

Abstract

Massive multiple-input multiple-output (mMIMO) downlink precoding offers high spectral efficiency but remains challenging to deploy in practice because near-optimal algorithms such as the weighted minimum mean squared error (WMMSE) are computationally expensive, and sensitive to SNR and channel-estimation quality, while existing deep learning (DL)-based solutions often lack robustness and require retraining for each deployment site. This paper proposes a plug-and-play precoder (PaPP), a DL framework with a backbone that can be trained for either fully digital (FDP) or hybrid beamforming (HBF) precoding and reused across sites, transmit-power levels, and with varying amounts of channel estimation error, avoiding the need to train a new model from scratch at each deployment. PaPP combines a high-capacity teacher and a compact student with a self-supervised loss that balances teacher imitation and normalized sum-rate, trained using meta-learning domain-generalization and transmit-power-aware input normalization. Numerical results on ray-tracing data from three unseen sites show that the PaPP FDP and HBF models both outperform conventional and deep learning baselines, after fine-tuning with a small set of local unlabeled samples. Across both architectures, PaPP achieves more than 21$\times$ reduction in modeled computation energy and maintains good performance under channel-estimation errors, making it a practical solution for energy-efficient mMIMO precoding.

Figures (4)

• Figure 1: Proposed "backbone" architecture. Only one student head (FDP or HBF) is activated at a time.
• Figure 2: Achievable sum rates for and methods under different SNR values at three Montreal sites: "Ericsson", "Decarie", and "Sainte-Catherine". All methods report zero-shot performance, except "PaPP + FT" and "PaPP Few-Shot", which are after 20 epochs of fine-tuning.
• Figure 3: Comparing achievable sum rates for and methods under different unseen deployment sites and SNR conditions. All methods report zero-shot performance, except "PaPP + FT" and "PaPP Few-Shot", which are after 20 epochs of fine-tuning.
• Figure 4: Impact of the channel estimation noise factor $\beta$ on the performance of the proposed PaPP models during deployment and fine-tuning for the "Ericsson" site at an average SNR of 20 dB.