EM-GANSim: Real-time and Accurate EM Simulation Using Conditional GANs for 3D Indoor Scenes

Ruichen Wang, Dinesh ManochaSubmission to ICLR 2025 · Rejected

TL;DR

EM-GANSim uses a physics-constrained conditional Generative Adversarial Network (cGAN) to predict 3D electromagnetic (EM) wave propagation in indoor environments in near real-time. By embedding electromagnetic theory directly into the loss function, the model achieves ray tracing-level fidelity while delivering a 5X overall speedup—computing point signal strength in just 1–2 milliseconds.

What Problem Does This Paper Address?

Designing indoor wireless networks (such as 5G and Wi-Fi) requires accurate modeling of how radio signals bounce, attenuate, and scatter off walls, doors, and furniture. Traditional simulation methods face a tough trade-off:

  • Full-wave numerical solvers (e.g., FDTD) are computationally prohibitive for large or interactive spaces.
  • Ray tracing (RT) is faster than full-wave modeling, but recalculating paths across complex or dynamic 3D layouts remains too slow for real-time applications, requiring dozens of seconds per room.

The authors address this bottleneck by designing a machine-learning surrogate model that produces accurate received-power heatmaps instantaneously.

Methods

EM-GANSim frames wireless signal prediction as an image-to-image translation problem conditioned on geometry and physics:

  • Conditional Architecture: The generator takes three inputs: an encoded spatial layout $E_g$ (capturing 2D footprint, heights, and material properties like concrete, wood, and glass), the 3D transmitter coordinates $L_{tx}$, and a Gaussian latent noise vector $z$. The discriminator evaluates pairs of geometry and power heatmaps to enforce visual and spatial consistency.
  • Physics-Constrained Loss ($\mathcal{L}_{phy}$): Rather than treating the task as purely visual regression, the generator's objective function balances standard adversarial loss and mean squared error (MSE) with an explicit propagation loss: $$\mathcal{L}_{phy} = \alpha \mathcal{L}_{direct} + \beta \mathcal{L}_{reflection} + \gamma \mathcal{L}_{diffraction}$$
  • Direct Path Loss: Penalizes deviation from free-space path loss (FSPL), distance attenuation, and atmospheric absorption.
  • Reflection Loss: Enforces energy drops calculated using material reflection coefficients and surface geometry.
  • Diffraction Loss: Employs a modified Uniform Theory of Diffraction (UTD) model to capture shadow boundaries around sharp edges and obstacles.

Datasets

The authors curated a dataset built from 2,000+ 3D indoor scenes and over 64 million simulated heatmaps:

  • Scenes: Ranging from $4\text{ m}^2$ single rooms to $144\text{ m}^2$ complex floor plans featuring varying partitions and combinations of concrete, wood, and glass.
  • Simulation Ground Truth: Generated using high-fidelity ray-tracing simulators—specifically WinProp (an industry standard) and DCEM.

Key Insights and Ideas

  • Physics constraints prevent visual hallucination: Standard CNN/GAN generators often produce smoothed or physically impossible energy distributions. Explicitly penalizing violations of UTD diffraction and Fresnel reflections forces the network to preserve realistic shadow boundaries and multipath fringes.
  • Latent noise acts as an essential regularizer: The ablation studies demonstrate that retaining Gaussian noise input $z$ is critical. Without noise, the network suffers from mode collapse and over-smoothing (e.g., MSE rose from $3.65\text{ dBm}^2$ to $16.38\text{ dBm}^2$ in Scene 13), whereas noise enables the network to learn rich multipath variations.
  • Global inference in one forward pass: Once trained, the network bypasses iterative ray-triangle intersection tests entirely, yielding the complete field distribution in a single neural network forward pass.

Key Results

  • Real-Time Speed: EM-GANSim generates full-room power heatmaps in 3.0 to 4.3 seconds, delivering an approximate 5X speedup over traditional ray-tracing tools (which required 12 to 22 seconds). Individual receiver points are evaluated in 0.9 to 2.0 milliseconds.
  • High Predictive Fidelity: On unseen test layouts, EM-GANSim achieves an average MSE of roughly $8.5\text{ dBm}^2$ relative to WinProp, closely tracking or matching dedicated ray tracers like DCEM.
  • Error Distribution: In statistical comparisons, EM-GANSim demonstrated tighter normalized error distributions around the median relative to baseline simulators, confirming reliable prediction even across complex multi-room geometries.

Limitations and Future Work

  • Ray Tracing Ground Truth Limitations: Because the model trains on synthetic ray-tracing outputs, it inherits ray tracing’s blind spots—such as low-frequency wave phenomena and complex diffuse scattering.
  • Static Environments: The current benchmark focuses on static indoor scenes; it does not yet model moving personnel or dynamic objects.
  • Indoor Scope: The pipeline is tailored to indoor environments; extending it to large-scale outdoor urban topologies with dense vehicle traffic remains future work. The authors plan to incorporate rich dynamic indoor benchmarks (such as 3D-FRONT) in subsequent iterations.

Summary updated

Abstract

We present a novel machine-learning (ML) approach (EM-GANSim) for real-time electromagnetic (EM) propagation that is used for wireless communication simulation in 3D indoor environments. Our approach uses a modified conditional Generative Adversarial Network (GAN) that incorporates encoded geometry and transmitter location while adhering to the electromagnetic propagation theory. The overall physically-inspired learning is able to predict the power distribution in 3D scenes, which is represented using heatmaps. Our overall accuracy is comparable to ray tracing-based EM simulation, as evidenced by lower mean squared error values. Furthermore, our GAN-based method drastically reduces the computation time, achieving a 5X speedup on complex benchmarks. In practice, it can compute the signal strength in a few milliseconds on any location in 3D indoor environments. We also present a large dataset of 3D models and EM ray tracing-simulated heatmaps. To the best of our knowledge, EM-GANSim is the first real-time algorithm for EM simulation in complex 3D indoor environments. We plan to release the code and the dataset.
BibTeX
@misc{wang2024em,
  author = {Ruichen Wang and Dinesh Manocha},
  title = {{EM-GANSim: Real-time and Accurate EM Simulation Using Conditional GANs for 3D Indoor Scenes}},
  year = {2024},
  eprint = {2405.17366},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.17366},
}

Loading related content...

Figure