Photon reconstruction with neural networks

This repository hosts the code for my bachelor thesis "Neural Networks for Photon Reconstruction in Electromagnetic Calorimeters for the COMPASS and AMBER Experiments" submitted on the 07.07.2023 at the Technical University of Munich.

Example of a two-photon data sample. This work aims to precisely determine the position (red crossed) and energies (measured by the individual cells) of each photon.

Abstract of the thesis

As traditional methods for photon reconstruction in electromagnetic calorimeters at the COMPASS and AMBER experiments at CERN are pushed to their limits, we consider an alternative approach using artificial neural networks. This thesis is a proof of principle study on whether neural networks can improve the reconstruction of photon positions and energies detected with the calorimeter on simulated data. The network can determine the position of a single photon with higher precision than the traditional fit method and shows roughly the same performance in energy reconstruction. Although the network was trained on simulated data, it succeeds in reconstructing real-data photons with the same accuracy as the traditional fit method. In the case of two overlapping photons, the network still performs better in position determination than the traditional method, however the energy prediction of the traditional method remains superior. This lack of performance might be due to the network’s architecture being chosen too simply. Examination of the network’s ability to count the amount of photons detected on the calorimeter showed that using neural networks might improve the separation of small-distance photon showers but should be trained sensibly to predict the correct number of photons in all cases. In short, artificial neural networks show potential in photon reconstruction in electromagnetic calorimeters but the network’s complexity and training process must be chosen carefully to achieve better results than traditional methods.

Structure of the repository

To approach the task of reconstructing the position ($x$, $y$) and energy $E$ of photons hitting the ECAL, we build up the complexity step by step.

1. Stage 1: Single, monoenergetic photons($E=1404 GeV)
2. Stage 2: Single photons ($E=<200$ GeV)
3. Stage 3: Single photons coming from different angles
4. Stage 4: Two photons, reconstruction locally (9x9 cells)
5. Stage 4, fullrange: Two photons
6. Stage 5: Counting the total amount of photons