Study effect of material and non-homogeneous magnetic field about sms2021-tra-tra HOT 2 OPEN

Most code is ready, I already made a pull request for it. The results take long to compute because of the evaluation metrics. I will leave it running overnight and I will post the plots in the morning, in case something doesn't break.

from sms2021-tra-tra.

For the initial analysis, I picked randomly an event from each dataset. The hits in the xy and rz space look like this:

The respective tracks in the Hough space look like this:

So the reason this has been going slow, is because with the following hyperparameters:

I have been getting these results, averaged for the whole datasets (for each type of transformation):

If I am not mistaken, this is not likely a correct result. As Noemi told me, the efficiency should drop to 0.6 for the simulation with material effects and non-homogenous B.

I am still looking for the bug. These are the commands I used to produce the datasets: https://github.com/asalzburger/tra-tra/blob/main/src/notebooks/issue8/production.cmd.

In the meanwhile, I have started developing a way to filter out duplicate tracks:

1. Scan the accumulator array and identify "hot" regions, i.e. with many filled bins close to each other. If those bins have the same tracks, then they are likely duplicates. We could also take this one step further and look at geometries: making sure that the hits are touching different layers. This can also work as a way to get rid of fakes (which is not a big issue in the combined transformation).

2. Use a Neural Network to decide if two bins are referring to the same track. We can have a fixed-size input layer that will take the xyz coordinates of every hit inside each bin, and will classify the two tracks as duplicates or not. The creation of a dataset for this part is a bit tricky, as we must somehow make sure that the data comes from the same distribution, and that also there is a balance.

Both ideas are still being developed.

from sms2021-tra-tra.