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.
Introduction
Designing modern wireless networks, such as Wi-Fi 6 or 5G, necessitates precise Electromagnetic (EM) simulations to predict signal behavior—including reflections off walls, diffraction around obstacles, and path loss. Traditional methods face a significant trade-off: Maxwell’s Equations solvers are highly accurate but computationally prohibitive for real-time applications, while Ray Tracing, though faster, remains too slow for dynamic environments or large-scale optimization tasks where thousands of transmitter placements must be evaluated instantly.
Purely data-driven AI models often fail in this domain because they can "hallucinate" signal patterns that violate fundamental physical laws. EM-GANSim addresses this by integrating physical constraints directly into a deep learning framework, achieving the speed of neural networks without sacrificing the reliability of physics-based engines.
Methodology
The core of EM-GANSim is a modified conditional Generative Adversarial Network (cGAN) architecture tailored for mapping 3D geometry and transmitter (Tx) locations to received power heatmaps.
Architecture
The system consists of two primary components:
- Generator: A convolutional neural network that processes encoded 3D geometry (spatial layout and material properties), the precise Tx location, and a Gaussian noise vector. The noise vector is crucial for avoiding "mode collapse" and capturing complex interference patterns.
- Discriminator: This component evaluates the generated heatmaps for realism against ray-tracing ground truth and ensures consistency with the input geometry.
Physics-Informed Loss Function
The innovation lies in the Physics-informed Loss Function ($L_{phy}$). Beyond standard pixel-wise differences, the model is penalized for violating EM theory. It explicitly accounts for:
- Direct Propagation: Enforcing the inverse-square law for path loss.
- Reflection: Modeling signal bounces based on material constants (e.g., concrete vs. glass).
- Diffraction: Utilizing the Uniform Theory of Diffraction (UTD) to simulate wave bending around corners.
Dataset and Training
To train EM-GANSim, the researchers developed a comprehensive dataset comprising over 2,000 3D indoor models with diverse layouts and materials. This resulted in more than 64 million data points. Ground truth heatmaps were generated using industry-standard simulators like WinProp and DCEM. The model was evaluated on 15 complex benchmarks and 4 "unseen" scenarios to verify its generalization capabilities.
Results and Discussion
EM-GANSim demonstrates a significant performance leap, achieving a 5X speedup over traditional ray tracing. It can predict signal strength at any location in approximately 1 millisecond.
Quantitatively, the Mean Squared Error (MSE) is comparable to traditional simulators. Qualitatively, as shown in the heatmap comparisons, the model accurately captures complex phenomena like shadowing behind walls and signal reinforcement from reflections. Crucially, the model generalizes well to entirely new floor plans, indicating it has learned the underlying principles of spatial propagation rather than just memorizing training maps.
Conclusion
EM-GANSim represents a major advancement in real-time EM simulation. By embedding physical constraints into a GAN framework, it provides a fast, accurate, and physically consistent tool for wireless network planning. Future work will focus on extending the model to high-density outdoor urban environments and incorporating dynamic elements like moving obstacles.
Summary updated