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

TL;DR

EM-GANSim presents a physics-informed conditional GAN to predict 3D indoor EM heatmaps in real time, conditioning on encoded geometry and transmitter location. The generator uses CNNs and a composite loss that blends adversarial, MSE, and physics terms for direct, reflected, and diffracted propagation, while the discriminator enforces geometry-consistent realism. On a large corpus of indoor scenes, EM-GANSim achieves real-time inference with roughly a 5x speedup over ray-tracing baselines, with accuracy approaching traditional RT methods and robust generalization across diverse environments. The work provides a large EM propagation dataset and outlines a path toward dynamic indoor and urban wireless simulations for rapid network planning and real-time decision support.

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. We evaluated our method on 15 complex 3D indoor environments, with 4 additional scenarios later included in the results, showcasing the generalizability of the model across diverse conditions. 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.

Figures (9)

• Figure 1: Overall architecture of our cGAN training process. The Generator (G) takes encoded 3D geometry, transmitter location, and a noise vector to output simulated heatmaps. The Discriminator (D) evaluates both the real heatmap from a ray-tracing simulator DCEM and the generated heatmap from G and makes 0/1 decisions.
• Figure 2: Representative examples of the indoor-scene geometry used in our experiments. We show three canonical layouts: (a) Single-room setup with minimal furniture. (b) Multi-room configuration with complex wall structures. (c) Multi-room layout with varied dimensions and partitions, drawn from our full dataset (> 2 000 models). These scenes demonstrate the diversity of layouts the ML model must interpret for accurate EM ray tracing simulation. The red represents concert walls, the blue represents glass, and the yellow represents wooden doors. These images serve purely as illustrative samples; quantitative evaluations are reported on the 15 (baseline) + 4 (additional) benchmarks detailed in Tables II–V and Figs 4–9.
• Figure 3: A more detailed flowchart of the GAN training process and implementation details: After data preparation, we encode geometry info along with transmitter location and a noise vector to feed into the generator networks. The generator employs a series of convolutional neural network (CNN) layers designed to capture the intricate spatial relationships within the indoor environments. Special attention is given to geometry information, allowing the model to understand how different materials and layouts affect signal propagation. The discriminator is also based on CNNs, with the addition of condition layers that incorporate geometry information. This setup ensures that the discrimination process considers not just the realism of the heatmaps but also their consistency with the input geometry. The loss function is selected as binary cross-entropy, backpropagated through the respective networks to compute the gradient of the loss with respect to the network weights. Gradient descent optimization algorithms are used to adjust the weights of the generator and discriminator in the direction that will reduce their respective losses.
• Figure 4: Comparative heatmaps displaying received powers in indoor environments of size 5*5 $m^2$ (left three columns, Scene 1-3) and 12*12 $m^2$ (right three columns, Scene 4-6). First row: WinProp simulation. Second row: GAN-based simulation. Third row: DCEM simulations. The MSEs of GAN-based and DCEM compared to WinProp are shown in Table \ref{['table:MSE1']} below. We see with GAN-based methods that the heatmaps show less MSE in general captures and exhibit more pronounced areas of both high and low signal strength, suggesting a finer granularity in the simulation of received powers.
• Figure 5: Comparative heatmaps displaying received powers in indoor environments. First row: WinProp simulation. Second row: GAN-based simulation. Third row: DCEM simulations. The room sizes on the right (Scene 10-12) are larger than those on the left (Scene 7-9).