Electromagnetic Raytracing1 has become much more relevant for future digital twin modelling, as it now supports differentiation of the raytracing output. Not only does this allow smooth interpolation of the raytracing result, for instance in space and time, but also the integration of raytracing into learning architectures, which in turn solve inverse problems. This term is coined physics based deep learning2.
At EMS we wish to first measure, then estimate, and finally model radio channels as accurately as possible. For this we develop dedicated, high precision measurement devices, so called channel sounders, that are able to capture spectral, spatial and temporal information about radio wave propagation. However, the effort in terms of hardware and the subsequent data processing is rising disproportionally with measurement bandwidth, scenario complexity and required accuracy.
To avoid this, we wish to pair the conducted radio measurements with a suitable parametric model that allow to accurately simulate the dominating propagation phenomena. This would allow us to generate radio channel realizations cheaply, quickly, especially in complex scenarios without costly measurements. In our case, so-called differentiable ray tracing tools can turn a digital description of an environment or an object into a simulation. This can be done at rapid speed due to GPU-based hardware acceleration.
