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

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.

💡 Research Summary

The paper tackles the long‑standing challenge of downlink precoding in massive MIMO (mMIMO) systems, where the optimal sum‑rate maximization problem is non‑convex and NP‑hard. Classical iterative algorithms such as the weighted minimum mean‑squared error (WMMSE) achieve near‑optimal performance but require cubic‑order matrix inversions at each iteration, making real‑time deployment on power‑constrained hardware impractical. Recent deep‑learning (DL) approaches replace the iterative solver with a direct CSI‑to‑precoder mapping, dramatically reducing inference latency. However, these models typically suffer from poor generalization: they must be retrained for each new deployment site, transmit‑power level, or channel‑estimation quality, limiting their practical usefulness.

To address these shortcomings, the authors propose PaPP (Plug‑and‑Play Precoder), a low‑complexity DL framework that can be trained once (the “backbone”) and then deployed across a wide variety of scenarios without full retraining. PaPP is designed for both fully‑digital precoding (FDP) and hybrid analog‑digital beamforming (HBF). Its key innovations are:

1. Teacher‑Student Architecture – A high‑capacity “teacher” network learns to emulate the WMMSE solution by predicting the auxiliary variables (weights v, receiver gains u, and Lagrange multiplier μ) in a numerically stable numerator/denominator form. A lightweight “student” network distills this knowledge and directly outputs the final precoding matrix (FDP) or the analog and digital components (HBF) without performing any matrix inversion. This dramatically reduces FLOPs while preserving the performance of the teacher.

2. Meta‑Learning Domain Generalization (MLDG) – During training, the backbone is exposed to multiple “domains” that differ in site geometry, SNR, and channel‑estimation error. By alternating between inner‑loop updates (within a domain) and outer‑loop updates (across domains), the model learns parameters that are robust to domain shifts, enabling rapid adaptation to unseen environments.

3. Transmit‑Power‑Aware Input Normalization – The raw CSI matrix is scaled by the transmit‑power level before being fed to the network. Consequently, the model implicitly learns power‑independent representations, eliminating the need for an explicit power input.

4. Dual Feature Extraction – The backbone processes the complex CSI through two parallel branches:

   • A three‑layer convolutional network extracts global spatial features across antennas.
   • A per‑user multi‑layer perceptron (MLP) processes the real‑imaginary concatenated channel vector for each user. User embeddings are summarized by mean and three fixed quantiles, producing a global context vector that is gated and merged back into the per‑user embeddings. The combined features are shared between teacher and student heads.

5. Self‑Supervised Loss – Training optimizes a weighted sum of (i) a teacher‑imitation loss (L2 distance between student and teacher outputs) and (ii) a normalized sum‑rate loss that directly encourages the student’s precoder to achieve a target fraction of the maximum achievable rate. This dual objective ensures that the student not only mimics the teacher but also maximizes actual system performance.

Experimental Evaluation
The authors evaluate PaPP on a realistic ray‑tracing dataset generated from detailed 3‑D maps of Montreal. Three unseen sites (industrial campus, dense downtown, suburban area) are used for testing. System parameters include 128 transmit antennas and 8–16 single‑antenna users. The backbone is pre‑trained on a large collection of channels and then fine‑tuned on each new site using only a few hundred unlabeled CSI samples (no ground‑truth precoders required).

Results show:

• Spectral Efficiency – Both PaPP‑FDP and PaPP‑HBF achieve sum‑rates that are 2–5 % higher than WMMSE on average and 5–12 % higher than the best existing DL baselines, across an SNR range of 0–30 dB.
• Robustness to CSI Errors – When the normalized mean‑square error (NMSE) of the channel estimate varies from –10 dB to –5 dB, the performance degradation is less than 2 %, demonstrating strong resilience to noisy or quantized CSI.
• Computational Savings – The student network eliminates matrix inversions, reducing FLOPs to roughly 4 % of WMMSE’s cost. Measured inference power drops from ~1 W (WMMSE) to ~0.05 W, corresponding to a >21× reduction in modeled computation energy.
• Generalization – Because the backbone is trained with MLDG and power‑aware scaling, the same model can be deployed across sites with different antenna layouts, user distributions, and transmit‑power budgets without any architectural changes.

Implications and Limitations
PaPP bridges the gap between high‑performance but computationally heavy optimization methods and fast but brittle DL models. Its plug‑and‑play nature makes it attractive for real‑world 5G/6G deployments where base stations may be upgraded or relocated frequently. The lightweight student can be implemented on ASICs or FPGAs, enabling ultra‑low‑latency, energy‑efficient precoding.

However, the current work focuses on fully connected hybrid architectures with a single RF chain per antenna (i.e., constant‑modulus phase shifters) and does not explore more restrictive hardware constraints such as limited phase‑resolution or sub‑array structures. Extending PaPP to those scenarios, as well as integrating it with channel‑feedback reduction techniques for FDD systems, are promising directions for future research.

In summary, the paper presents a novel, generalizable, and energy‑efficient deep‑learning precoding framework that achieves near‑optimal sum‑rate performance while drastically cutting computational overhead, marking a significant step toward practical massive MIMO deployments.