Physics-Aware Conditional SetGAN for Spatially Consistent Multi-User TR 38.901 Channel Generation

Technical Summary: Physics-Aware Conditional SetGAN for Spatially Consistent Multi-User TR 38.901 Channel Generation

Problem Statement
Generating multi-user multiple-input multiple-output (MIMO) channel realizations consistent with the 3GPP TR 38.901 standard is computationally expensive, particularly when evaluating large numbers of user equipment (UE) deployments. While tools like Sionna provide reliable, physics-grounded references, their runtime cost becomes prohibitive for extensive Monte Carlo simulations. A critical challenge is that multi-user channels cannot be treated as independent draws; nearby UEs must exhibit correlated large-scale effects and coherent small-scale fading structures to maintain spatial consistency. Existing generative models often focus on individual links, global statistics, or specific tasks, failing to reproduce the full geometry-conditioned joint structure of multi-user snapshots with explicit spatial consistency.

Methodology
The authors propose a Physics-Aware Conditional SetGAN designed to generate multi-user channel realizations faster than the reference simulator while preserving spatial correlations imposed by user geometry. The approach follows an end-to-end workflow comprising five stages:

1. Data Representation and Factorization:

   • Reference data is generated using Sionna for the UMa/NLoS scenario.
   • The channel tensor is decomposed into two components: a large-scale received power term (slowly varying) and a normalized small-scale fading tensor (fast fluctuations).
   • The large-scale power is kept as a direct scalar prediction target.
   • The normalized small-scale fading is compressed using Principal Component Analysis (PCA) into a whitened latent space to reduce dimensionality and facilitate adversarial learning.

2. Architecture (Geometry-Conditioned SetGAN):

   • Set-Based Formulation: The model treats the multi-user snapshot as an unordered set of UEs, ensuring permutation invariance/equivariance.
   • Geometry Awareness: The generator utilizes Set Transformers with relative positional biases. User terminal coordinates are encoded via random Fourier features and passed through a learned MLP.
   • Interaction Mechanism: The attention mechanism includes a learned pairwise bias term dependent on the distance between user terminals ($b(r_i, r_j)$). This imposes a distance-aware interaction prior, encouraging stronger coupling between nearby UEs to mimic physical spatial consistency.
   • Noise Injection: The model accepts global noise (shared across the snapshot) and local noise (specific to each UE) to capture both collective and individual channel variations.

3. Physics-Informed Training:

   • The training objective combines a standard adversarial loss (Wasserstein GAN with spectral normalization) with specific physics-aware auxiliary losses.
   • Profile Matching: The model explicitly minimizes the error between generated and reference similarity profiles (fading-structure and received-power) as a function of UE separation distance.
   • Tail Regularization: Specific terms ($L_{topk}$, $L_{mass}$) are introduced to ensure fidelity in the tails of the received-power distribution, which standard adversarial terms often fail to capture accurately.
   • Supervision: Direct regression losses are applied to the large-scale received power and smoothness constraints are applied to local power variations.

Key Results
Evaluated on the UMa/NLoS benchmark with 100 UEs per snapshot, the trained surrogate demonstrates the following performance relative to the Sionna reference:

• Fidelity:
  • Received Power: The generated power distribution closely matches the reference, with a Wasserstein-1 distance of 0.41 dB and a Kolmogorov-Smirnov statistic of 0.0170.
  • Spatial Consistency: The model reproduces spatial-consistency profiles with mean absolute errors (MAE) on median curves below 0.03 (specifically 0.0255 for power similarity and 0.0260 for fading-structure similarity) across distance bins.
• Efficiency:
  • In a fixed-position CPU-vs-CPU benchmark, the model reduces the elapsed generation time by a factor of 3.45 and the total CPU cost by a factor of 6.15 compared to Sionna.
  • In a GPU-vs-GPU benchmark (with 16 GB memory caps), speedups are 2.52× for generation time and 1.31× for total cost.

Significance and Claims
The paper claims that a trained generative model can substantially accelerate TR 38.901 channel generation without compromising the spatial consistency required for evaluating multi-user systems. The primary contribution is a surrogate that combines:

1. Geometry-conditioned generation of full multi-user snapshots.
2. Explicit preservation of spatial consistency through distance-profile matching.
3. Direct fidelity and runtime benchmarking against a matched-position Sionna reference.

The authors emphasize that this work resolves the specific gap of generating full geometry-conditioned TR 38.901 snapshots with explicit spatial-consistency profiles, rather than just individual links or statistical covariances. The results are validated under a controlled UMa/NLoS protocol; the authors note that broader claims across other TR 38.901 scenarios or deployment settings require further validation.