⚡ electrical engineering

Site-Specific Beam Learning for Full-Duplex Massive MIMO Wireless Systems

This paper proposes a novel deep learning-based beam learning framework for full-duplex massive MIMO systems that eliminates the need for explicit self-interference channel estimation, achieving high signal-to-noise ratios with 75-97% fewer measurements than traditional methods.

Original authors: Samuel Li, Ian P. Roberts

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samuel Li, Ian P. Roberts

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to have a conversation with a friend while standing right next to a massive, roaring jet engine. In the world of wireless communication, this is exactly what a Full-Duplex system tries to do: a cell tower talks to your phone (sending data) and listens to your phone (receiving data) at the exact same time on the exact same frequency.

The problem? The tower's own voice is so loud that it drowns out the whisper from your phone. This is called Self-Interference.

The Old Way: The "Blindfolded Detective"

Traditionally, to solve this, the tower had to act like a detective trying to map the echo of its own voice. It would shout a specific test signal, listen to the echo, and calculate exactly how the sound bounces off the walls, the ground, and the air to figure out how to cancel it out.

In modern "Massive MIMO" towers (which have hundreds of tiny antennas), this detective work is a nightmare. To map the echo perfectly, the tower would need to shout thousands of different test signals. This takes up so much time and energy that there's no time left to actually send data. It's like spending all day measuring the room's acoustics and never actually having the conversation.

The New Way: "Site-Specific Beam Learning"

The researchers at UCLA (Samuel Li and Ian Roberts) came up with a clever shortcut. Instead of trying to map the entire echo chamber perfectly, they taught the tower to learn the shape of the room using a few quick "sniffs" and then use a smart brain (AI) to figure out the best way to talk and listen.

Here is how their new system works, broken down into simple steps:

1. The "Sniff" (Probing)

Instead of shouting thousands of test signals, the tower uses a pre-designed set of just a few special "sniffing" beams (let's say 8 to 64 of them).

Analogy: Imagine you are in a dark room with a bat. Instead of trying to map every single inch of the room, you just make a few specific clicks. You don't need to know the exact distance to every dust mote; you just need to know the general "shape" of the echo so you know where not to shout.

2. The "Smart Brain" (Deep Learning)

The tower feeds these few "sniff" results into a neural network (a type of AI). This AI has been trained on the specific building and street where the tower is located.

Analogy: Think of this AI as a local guide who has lived in that specific neighborhood for years. Even if the guide only sees a few landmarks (the "sniffs"), they know exactly where the traffic jams (interference) are and how to route the cars (data) to avoid them. The AI doesn't need a perfect map; it just needs enough clues to make a smart guess.

3. The "Perfect Dance" (Beam Synthesis)

Based on those few clues, the AI instantly designs the perfect "dance moves" for the antennas. It tells the transmitting antennas how to aim their signal away from the receiving antennas, and tells the receiving antennas how to listen away from the transmitter.

Analogy: It's like two dancers who know exactly how to move so they never bump into each other, even though they are spinning in a tiny space. They don't need to measure the floor; they just know the rhythm.

Why is this a Big Deal?

The paper shows that this new method is incredibly efficient:

Speed: It requires 75% to 97% fewer measurements than the old way. It's the difference between spending an hour measuring a room versus just glancing at it for a second.
Performance: It successfully cancels out the "roaring jet engine" noise, allowing the tower to talk and listen simultaneously with very high clarity.
Adaptability: Because the AI is trained on the specific "site" (the specific building and street), it learns the unique quirks of that location. If the wind blows or a truck drives by, the AI is already tuned to handle the echoes of that specific place.

The Bottom Line

This paper proposes a shift from "measuring everything perfectly" to "learning just enough to be smart." By using a little bit of targeted probing and a powerful AI brain, cell towers can finally talk and listen at the same time without getting deafened by their own voices. This paves the way for faster, more efficient 6G networks that can handle massive amounts of data without slowing down.

1. Problem Statement

The paper addresses the critical challenge of implementing in-band full-duplex (FD) operation in millimeter-wave (mmWave) and massive MIMO systems.

The Core Issue: Full-duplex operation (transmitting and receiving simultaneously on the same frequency) is hindered by severe self-interference (SI) at the base station (BS).
The Bottleneck: Existing solutions rely on explicit estimation of the SI channel ( $H$ $H$ ). In massive MIMO systems with $N_t$ $N_{t}$ transmit and $N_r$ $N_{r}$ receive antennas, the SI channel matrix size is $N_r \times N_t$ $N_{r} \times N_{t}$ . Explicitly estimating this channel requires $N_t N_r$ $N_{t} N_{r}$ pilot measurements.
- For a typical 16-antenna array, this requires 256 measurements.
- In fast-fading environments or high-mobility scenarios, this pilot overhead is prohibitively high, rendering explicit estimation impractical for 5G/6G networks.
Limitations of Current Alternatives: Previous attempts to avoid explicit estimation (e.g., exhaustive grid searches or LSTM-based sequential probing) suffer from high overhead per user or poor scalability in multi-user scenarios.

2. Methodology

The authors propose a novel Site-Specific Beam Learning Framework that bypasses explicit channel estimation. Instead of estimating the full channel matrix, the system learns to synthesize optimal beams using implicit channel knowledge derived from a small number of probing measurements.

A. System Model

Setup: A BS with separate transmit ( $N_t$ ) and receive ( $N_r$ ) arrays serves a downlink (DL) user and an uplink (UL) user simultaneously.
Objective: Maximize the sum spectral efficiency ( $R_{DL} + R_{UL}$ ) by designing transmit beam $f$ and receive beam $w$ .
Constraints: The design must adhere to per-antenna power constraints and minimize the interference-to-noise ratio (INR) caused by self-interference.

B. The Proposed Framework

The solution consists of three main components trained end-to-end:

Probing Codebooks ( $F, W$ ):
- Instead of estimating the full $H$ , the BS transmits/receives using a set of $M$ probing beam pairs $(f_m, w_m)$ , where $M \ll N_t N_r$ .
- These beams collect scalar measurements $z_m$ representing the projection of the SI channel onto the beam directions.
- The codebooks are site-specific, meaning they are optimized for the specific physical environment (geometry, reflections) of the deployment site.
Deep Learning Beam Synthesizer ( $D(\cdot)$ ):
- A neural network takes as input:
  - The implicit SI measurements ( $z$ ).
  - The DL and UL channel knowledge ( $y_{DL}, y_{UL}$ ), obtained via standard beam alignment.
- The network outputs the optimal serving beams $(f^*, w^*)$ .
Joint Optimization:
- The probing codebooks and the neural network are trained jointly to maximize the expected sum spectral efficiency.
- Training Constraint: To ensure physical realizability (unit modulus constraints on phase shifters), the real and imaginary parts of the codebook weights are updated separately and projected onto the feasible set after each gradient step.

C. Operational Workflow

Initialization: Acquire DL/UL channel knowledge (e.g., via beam alignment).
Probing: Execute $M$ probing measurements using the learned codebooks to gather implicit SI data.
Inference: Feed the measurements and channel data into the trained neural network to generate the final serving beams.
Reuse: Since the SI channel is often more stable than user channels (or varies on a different timescale), the same probing measurements can be reused for multiple user pairs, drastically reducing overhead.

3. Key Contributions

Implicit Estimation via Deep Learning: The paper introduces a framework that replaces explicit $N_t N_r$ channel estimation with a deep learning model trained on $M$ implicit measurements ( $M \ll N_t N_r$ ).
Site-Specific Codebook Design: Unlike generic codebooks, the probing beams are optimized specifically for the deployment site's channel characteristics, allowing the system to "learn" the environment's interference patterns.
User-Agnostic Overhead Reduction: The probing phase is independent of the specific user pair. Once the SI knowledge is gathered, it can be applied to any user pair, making the solution scalable.
End-to-End Trainability: The integration of trainable codebooks and a neural network synthesizer allows the system to adapt to complex, non-linear interference environments that traditional linear estimation cannot handle efficiently.

4. Numerical Results

The authors evaluated the framework using ray-tracing simulations of a 28 GHz base station (16 antennas) deployed at UCLA (Royce Hall/Dickson Plaza).

Measurement Efficiency:
- Explicit estimation requires 256 measurements (for 16x16 MIMO).
- The proposed method achieves near-optimal performance with only 8 to 64 measurements.
- This represents a 75–97% reduction in pilot overhead compared to explicit estimation.
Self-Interference Suppression:
- The baseline (MRT+MRC without SI cancellation) suffers from ~15 dB INR, making full-duplex infeasible.
- The proposed method with just $M=16$ probing beams reduces INR by nearly 20 dB, pushing self-interference below the noise floor in most cases.
Robustness to Channel Conditions:
- Rician Factor ( $\kappa$ ): Performance improves as the LOS component of SI increases (more deterministic), but the model remains effective even in stochastic (low $\kappa$ ) environments.
- NLOS Reflections: The method adapts well to varying numbers of reflections. While performance degrades slightly in highly complex Rayleigh fading environments with very few measurements, it significantly outperforms baselines.
Spectral Efficiency: The system achieves sum spectral efficiencies approaching the theoretical full-duplex capacity (which assumes perfect SI cancellation), validating the effectiveness of the learned beams.

5. Significance

This work is significant for the advancement of 6G wireless networks for several reasons:

Practicality of Full-Duplex: It removes the primary barrier to mmWave full-duplex deployment—the prohibitive pilot overhead required for SI estimation.
Scalability: By decoupling the SI probing from the user data, the solution scales efficiently to massive MIMO arrays and multi-user scenarios.
Data-Driven Physical Layer: It demonstrates the power of combining site-specific hardware configuration (codebooks) with deep learning to solve complex physical layer optimization problems that are difficult to model analytically.
Future Direction: It paves the way for "learning-based" wireless systems where the network learns the physics of its specific environment to optimize performance dynamically, rather than relying on static, worst-case assumptions.