Master projects in the ATLAS group
The Department of Particle Physics (DPNC) at University of Geneva is looking for master students interested in pursuing a
Master project in experimental particles physics
within the ATLAS group.
The DPNC is strongly involved in the operation of the ATLAS experiment. It is contributing to event reconstruction and searches for new physics, as well as the upcoming High-Luminosity LHC (HL-LHC) upgrade. As part of our on-going effort to investigate the use of novel machine learning techniques in the experimental particle physics community, we are developing new approaches to particle track reconstruction at ATLAS based on different data science techniques.
Possible topics for a master project are:
Similarity hashing-based reconstruction chain evaluationBased on the similarity hashing technique a full track reconstruction chain can be build. This comprises a pre-selection of possible hits followed by hit clustering into tracks within a pre-selected bucket. With the full chain, we can perform a detailed study of the performance as a function of the relevant kinematic parameters, e.g. particle type, particle parameters, as well as the parameters of the hashing algorithms.
Similarity hashing-based reconstruction for long-lived particlesLong-lived particles with displaced vertices far from the interaction point are an interesting signature in the ATLAS experiment. Traditional track reconstruction does not perform well for these displaced track. Similarity hashing-based track reconstruction can potentially handle these tracks much better. A consistent performance evaluation for such topologies will be performed.
Track parameter estimation with neural networks
Track parameters are traditionally estimated using global least square or recursive Kalman filter techniques. These rely on excellent knowledge of the detector geometry, magnetic field, material descriptions, etc. . This means that they can consume a significant amount of computing power. Linearized and approximate versions exist, but often suffer from degraded performance especially if the underlying detector and field are inhomogeneous. For trigger applications it would be beneficial to have a fast, reasonable track parameter estimator for short, fixed-sized tracks. In this project, we will try to find an optimal neural network architecture to estimate such track parameters, investigate its performance, and try to optimize its interference performance.
Fast triplet-based track finding with neural networksFinding tracks from a set of measured hit positions is a crucial task during event reconstruction. In a combinatorial approach, one of the important task is to quickly decide if a set of three hits, a triplet, can potentially originate from the same particle. In this project, we will try to find an optimal triplet classifier based on a (deep) neural network, investigate possible input parameters, optimal architecture, and investigate its performance. Possible extensions to doublets or quadruplets of hits could also be considered.
Inquiries should be sent to