Geometric Wireless Simulation with Equivariant Transformers

Thomas Hehn, Markus Peschl, Tribhuvanesh Orekondy, Arash Behboodi, Johann Brehmer

Track: Extended abstract

Keywords: Differentiable simulation, inverse problems, learning to simulate, geometric deep learning, equivariance, wireless communication, electromagnetic signals

TL;DR: Neural surrogates for wireless signal propagation based on geometric inductive biases for inverse problems

Abstract:

Modelling the propagation of electromagnetic signals is critical for designing modern communication systems.
While there are precise simulators based on ray tracing, they do not lend themselves to solving inverse problems or the integration in an automated design loop.
We propose to address these challenges through differentiable neural surrogates that exploit the geometric aspects of the problem.
We introduce the Wireless Geometric Algebra Transformer (Wi-GATr), a generic E(3) equivariant backbone architecture for simulating wireless propagation in a 3D environment.
Further, we introduce two datasets of wireless signal propagation in indoor scenes.
On these datasets, we show the data-efficiency of our model on signal prediction and inverse problem solving capabilities using differentiable predictive modelling as well as diffusion models.

Submission Number: 36