# Hi Thanks for the introduction, I am Jérome Eertmans and I will present our work on comparing differentiable and dynamic ray tracing. This work was done in collaboration with the University of Bologna and UCLouvain. All the materials, including the slides, are available on GitHub or on my personal website, links at the end.
# Context We are interested with radio propagation modeling. As technologies evolve, the need for simulating dynamic scenes increases. Indeed, more and more applications assume an environment that is not static, where antennas and reflectors are moving.
## Context examples In this example, we have a street canyon with many cars that are moving.
When using an idiomatic Ray Tracing (RT) approach, we need to recompute paths for each scene variation, which become computationally expensive. While Ray Tracers and computers have become faster, recomputing all the paths everytime may not be the smartest approach.
Dummy slide after loop
Let us recall the basics of Ray Tracing. A scene can be representated at any time as... some TX, some RX and some objects.
We then trace the rays between TX and RX
And, using each reach ray, we compute the EM fields
E.g., EM fields can be used to compute the coverage map
If you look at the bigger picture, radio-wave propagation through RT is a two-step process.
## Dynamic scenes What happens when we change the scene? As just said, we could recompute the rays and EM fields, but this is expensive. If the scene only scene changes a little, a suggestion is that we could use the previous rays or EM fields, and local derivatives, to predict the new rays or EM fields.
We have two approaches
If we name the path tracing step as f(x).
And a full RT simulation as a 'snapshot'
Then we extrapolate future snapshots using previous ones and local derivatives
Diff. RT is about being able to differentiate any function of our code with respect to any parameter.
E.g., we define an objective function g(x), here, the power.
And use AD to perform gradient descend to find the optimal parameters.
We observed the following limitations:
Our contributions are as follows
Contents of this presentation
We rapidly compare the two approaches
We will use a simple example to illustrate the limits of snapshot extrapolation.
When performing RT, each reflection order is computed separately. E.g., the LOS path is computed first. Here, the lid region is the area where a LOS path exists.
Of course, ray tracing is not limited to first-order reflection.
By adding contributions from all previous reflections, we draw the multipath cells.
The superposition of all cells if what we call the Multipath Lifetime Map. A cell is defined as the area where the multipath structure remains the same, and a cell can be split into multiple regions.
And of course, the map changes when any other object, like TX, moves.
Dummy slide after loop
We define two metrics
In practice, how to we compute the MLM?
We place a grid of RXs in the scene and identify each path component
We then repeat the process for all RXs, identifying uniques cells. All the implementations details are available in the paper and in the provided tutorial.
We can then compute the metrics. If we know the density of RXs, the cell area can be estimated by counting the number of RX per cell.
We then compute the distance to the closest RX that is not in the same cell, and average the distance over all RXs in the cell to obtain the average inter-cell distance.
# MLM example
Animate MLMs
A discussed in the paper, metrics shows similar results to what is obtained from measurements of the coherence distance, time, stationarity, etc.
Let's wrap up
Finals words
Histogram
Table