Technical Summary: Site-Specific MIMO Channel Generation via Diffusion and Flow Matching

1. Problem Statement

Future wireless systems are envisioned as "AI-native," requiring high-quality, site-specific radio data throughout the model lifecycle for both training and evaluation. However, acquiring real-world site-specific channel data through measurement campaigns is time-consuming, costly, and often infeasible at scale. While generative models offer a solution to augment scarce datasets, existing approaches face challenges:

• Generative Adversarial Networks (GANs): Often suffer from training instability and mode collapse, with evaluations frequently limited to selected statistics rather than full high-dimensional channel realizations.
• Diffusion Models: Provide stable training and diverse sampling but typically require iterative denoising processes that are computationally expensive and introduce high latency during inference.
• Evaluation Gap: Many prior works lack comprehensive assessments that simultaneously address statistical fidelity, sampling efficiency, and utility in downstream physical-layer tasks.

The core problem addressed is the generation of high-fidelity, location-conditioned Multiple-Input Multiple-Output (MIMO) channel matrices that preserve the spatial structure of a deployment site, while balancing generation quality with inference efficiency.

2. Methodology

The paper proposes and compares two location-conditioned generative paradigms designed to learn the conditional distribution $p_\theta(\mathbf{H} | \mathbf{c})$, where $\mathbf{H}$ is the MIMO channel matrix and $\mathbf{c}$ represents the User Equipment (UE) coordinates.

A. Data and Representation

• Dataset: The study utilizes site-specific ray-tracing (RT) data generated via Sionna RT from a 3D dense urban scene in London. Scenarios include 28 GHz (Line-of-Sight, LoS) and 3.5 GHz (LoS and Non-Line-of-Sight, NLoS).
• Input Representation: Channels are transformed into a beamspace representation ($\mathbf{H}_v$) using 2D Fourier codebooks for both the Base Station (BS) and UE uniform planar arrays (UPA). The complex beamspace is converted into a real-valued tensor (separating real and imaginary parts) for model input.
• Conditioning: Both models are conditioned on UE coordinates ($\mathbf{c}$) and continuous generative time ($t$). Conditioning variables are injected into intermediate decoder feature representations via learned modulation layers (MLPs), allowing the network to adapt to spatial propagation conditions without altering the convolutional backbone.

B. Generative Architectures

Both models share a common convolutional backbone with residual connections and multi-scale feature aggregation (15.5M parameters).

1. Conditional Denoising Diffusion Implicit Model (cDDIM):

   • Mechanism: Learns a reverse denoising process. During training, it estimates the noise component added to a clean channel sample at time $t$.
   • Inference: Generates channels by iteratively removing noise from a Gaussian sample through a sequence of deterministic denoising steps.
   • Trade-off: High fidelity but requires multiple sampling steps, leading to higher latency.

2. Conditional Flow Matching Model (cFMM):

   • Mechanism: Learns a deterministic velocity field that defines a probability flow from a simple base distribution (Gaussian noise) to the target conditional distribution.
   • Inference: Generates channels by integrating an Ordinary Differential Equation (ODE) from $t=1$ to $t=0$.
   • Trade-off: Offers a lower-latency alternative, potentially requiring fewer numerical integration steps to achieve comparable quality.

3. Key Contributions

• First Comparative Evaluation: This work presents the first evaluation of both diffusion and flow-matching models for site-specific MIMO channel generation across three dimensions: statistical fidelity, generation efficiency, and downstream utility.
• Unified Framework: Both paradigms are implemented with identical channel representations, conditioning mechanisms, and neural backbones, ensuring that performance differences are attributable to the generative objectives and sampling procedures rather than architectural variations.
• Comprehensive Metrics: The study introduces a multi-faceted evaluation framework including:
  • Fidelity: Dominant beam consistency, beamspace power profile similarity (cosine similarity), and effective MIMO channel rank (Wasserstein distance).
  • Efficiency: Analysis of pre-training data efficiency (scaling with training set size) and inference sampling efficiency (fidelity-latency trade-off).
  • Downstream Utility: Assessment in Channel-State Information (CSI) compression and site-specific beam alignment tasks.
• Open Source: The complete training and evaluation pipeline is publicly released to facilitate reproducibility.

4. Results

A. Statistical Fidelity

• Both models accurately capture site-specific characteristics, preserving spatial consistency across different UE locations.
• Beam Consistency: Both models maintain high fidelity in dominant beam direction, even in challenging 3.5 GHz LoS+NLoS scenarios.
• Rank Distribution:
  • In LoS-only scenarios, cDDIM achieves lower Wasserstein distance for effective rank, better preserving the low-rank structure.
  • In mixed LoS+NLoS scenarios, cFMM outperforms cDDIM in rank distribution, suggesting it better captures the broader effective-rank distribution induced by rich multipath.
• Data Scarcity: Both models are effective in limited-data regimes (e.g., 200 training samples), capturing dominant beamspace structures with limited ground-truth data.

B. Generation Efficiency

• Inference Latency: cFMM demonstrates a significantly more favorable fidelity-latency trade-off. It achieves quality comparable to cDDIM with roughly one order of magnitude less inference time.
• Sampling Steps: cFMM reaches high cosine similarity (approx. 0.9) with only 10–15 ODE integration steps. In contrast, cDDIM requires a substantially larger number of denoising iterations (e.g., 150+) to reach similar fidelity levels.

C. Downstream Utility

• CSI Compression: Augmenting scarce datasets (e.g., 200 samples) with synthetic channels from either model significantly reduces Normalized Mean Square Error (NMSE) compared to using scarce data alone or augmenting with generic 3GPP stochastic channels. The generated channels approach the performance of a full 10k-sample ground-truth dataset.
• Beam Alignment: At 28 GHz, models trained with generated channels significantly outperform those trained on scarce data or 3GPP stochastic data.
  • cDDIM-based training approaches the performance of training on 10k real ground-truth channels (within ~1 dB).
  • cFMM achieves slightly lower but still competitive performance with much faster generation.

5. Significance and Claims

The paper claims that generative models, specifically conditional diffusion and flow matching, are viable instruments for relieving the bottleneck of data acquisition in AI-native wireless networks.

• Site-Specificity is Critical: The results demonstrate that simply increasing training data with generic stochastic channels (e.g., 3GPP) is insufficient. Synthetic channels must reflect site-specific propagation behaviors to improve downstream task performance.
• Efficiency vs. Fidelity: While diffusion models (cDDIM) offer high stability and fidelity, Flow Matching (cFMM) emerges as a particularly attractive candidate for practical deployment where low-latency channel synthesis is required, offering comparable quality with significantly reduced computational cost.
• Robustness: Both approaches are robust even when trained on scarce ground-truth data, making them suitable for scenarios where extensive measurement campaigns are infeasible.

The authors conclude that these generative models successfully learn physically meaningful MIMO characteristics rather than merely matching low-level statistics, thereby validating their utility for training and evaluating AI-driven physical layer functions.