FACE: Flow-Guided Angular Channel Estimation Framework for mmWave MIMO Systems

Samitha Gunarathne, Eslam Eldeeb, Nurul Huda Mahmood, Matti Latva-ahoarXiv 2026 · arXiv

TL;DR

The FACE (Flow-Guided Angular Channel Estimation) framework leverages Flow Matching—a cutting-edge generative AI technique—to accurately estimate high-dimensional mmWave channels. By exploiting angular-domain sparsity, FACE achieves superior accuracy and pilot efficiency compared to diffusion-based models while drastically reducing computational costs and inference time.

What Problem Does This Paper Address?

In the push toward 6G, millimeter-wave (mmWave) and Terahertz (THz) communications are essential for high data rates. These systems use massive MIMO (Multiple-Input Multiple-Output) antenna arrays to overcome high path loss. However, to work effectively, the system needs precise Channel State Information (CSI).

Acquiring CSI is notoriously difficult because:

  1. High Dimensionality: The number of antenna elements is huge, making the channel matrix massive.
  2. Limited Pilots: Sending too many training signals (pilots) eats up bandwidth.
  3. Low SNR: Before beamforming is established, the signal-to-noise ratio is often very low.

Conventional methods like Least Squares (LS) or Compressed Sensing (CS) often fall short due to high complexity or "basis mismatch" (where real-world signal angles don't perfectly align with a predefined grid). Recent diffusion-based generative models offer high accuracy but are computationally "heavy," requiring hundreds of iterative steps to produce an estimate.

Methods

The authors propose FACE, which stands for Flow-Guided Angular Channel Estimation. It introduces two key innovations:

  1. Angular-Domain Sparsity: Instead of working with the dense spatial channel matrix, FACE transforms the problem into the angular domain. Because mmWave signals typically travel along just a few dominant paths, the angular representation is "sparse"—mostly zeros with a few high-energy peaks. This structure is much easier for a generative model to learn.
Fig. 1: Normalized instantaneous channel power gain of the spatial- and angular-domain channels for L = 20: (a) spatial-domain channel H and (b) angular-domain channel Ha.
  1. Flow Matching (FM): This is the "secret sauce." Unlike diffusion models that learn to reverse a stochastic (random) noise process, Flow Matching learns a deterministic velocity field. This creates "straighter" paths from a simple noise distribution to the complex channel distribution. During inference, FACE uses a Plug-and-Play Proximal Gradient Descent (PnP-PGD) framework, alternating between ensuring the estimate matches the received pilots and "denoising" it using the learned flow matching prior.
Fig. 2: Proposed FACE framework for mmWave channel estimation.

Datasets

The framework was rigorously tested on two types of data:

  • Synthetic Dataset: Generated using a standard geometric channel model at 60 GHz, covering various antenna geometries (Uniform Linear Arrays and Uniform Planar Arrays) and varying numbers of signal paths.
  • Real Measurement Dataset: The authors used a publicly available 60 GHz channel sounding dataset collected in real-world environments. This ensures the model isn't just "overfitting" to a mathematical formula but can handle practical propagation conditions.

Key Insights and Ideas

  • Straight Paths are Faster: Because Flow Matching learns deterministic, near-straight trajectories, it requires significantly fewer sampling steps (e.g., 10–20 steps) compared to the 50–100+ steps often needed by diffusion models.
  • Sparsity as a Structural Prior: By training specifically on the angular domain, the model learns the physics of the channel—that it should be sparse—rather than just trying to denoise a random matrix.
  • Robustness to "Pilot Starvation": FACE maintains high accuracy even when the number of pilots is significantly lower than the number of antennas, a scenario where conventional estimators typically fail.

Key Results

  • Superior Accuracy: FACE consistently outperformed LMMSE, OMP, and diffusion-based estimators in terms of Normalized Mean Squared Error (NMSE) across all tested SNRs.
  • Efficiency Gains: FACE achieved target accuracy with roughly 10x to 20x fewer GFLOPs than diffusion-based methods. For instance, to reach -25 dB NMSE, FACE required ~20.5 GFLOPs compared to 205 GFLOPs for the diffusion baseline.
  • Near-Perfect Rates: The average achievable data rate using FACE-estimated channels was nearly identical to the "Perfect CSI" benchmark, proving its practical utility for beamforming.
NMSE performance vs. SNR for different antenna configurations. FACE (blue line) consistently stays below the baselines.

Limitations and Future Work

While FACE is highly effective, the authors noted that in real-world measurement data, the estimation error (NMSE) tends to saturate at high SNR. This suggests that there are measurement imperfections or hardware non-idealities in real systems that the current generative prior doesn't fully capture.

The authors suggest that future work could extend this framework to handle interference management in multi-user MIMO systems, which would be a critical step for deployment in dense 6G networks.

Summary updated

Abstract

Acquiring accurate channel state information (CSI) is essential for millimeter-wave (mmWave) multi-input multi-output (MIMO) systems. However, it remains challenging due to high-dimensional channels, limited pilot observations, and measurement noise. To address these challenges, we propose flow-guided angular channel estimation (FACE), a novel channel estimation framework based on flow matching, an emerging generative modeling paradigm that learns continuous transport dynamics between probability distributions. FACE exploits the inherent angular-domain sparsity of mmWave channels and incorporates the learned flow matching model as a generative channel prior. Numerical results show that FACE consistently achieves higher channel estimation accuracy than both conventional estimators and diffusion-based generative estimator across diverse channel and system configurations. Moreover, FACE maintains high estimation accuracy with fewer pilot observations and achieves an average achievable rate close to the perfect-CSI benchmark. Compared with diffusion-based estimation, FACE requires substantially fewer sampling steps while achieving superior or comparable estimation accuracy, reducing inference time and computational cost. Overall, the proposed framework demonstrates the potential of flow matching for accurate, pilot-efficient, and computationally efficient mmWave MIMO channel estimation.
BibTeX
@misc{gunarathne2026face,
  author = {Samitha Gunarathne and Eslam Eldeeb and Nurul Huda Mahmood and Matti Latva-aho},
  title = {{FACE: Flow-Guided Angular Channel Estimation Framework for mmWave MIMO Systems}},
  year = {2026},
  eprint = {2611.21698},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21698},
}
Figure