School of Physics - Theses
Permanent URI for this collection
Search Results
1 - 2 of 2
-
ItemAddressing domain shift in deeply-learned jet tagging at the LHC( 2023-09)Over the last fifteen years, deep learning has emerged as an extremely powerful tool for exploiting large datasets. At the Large Hadron Collider, which has been in operation over the same time span, an important use case is to identify the initiating particles of hadronic jets. Due to the complexity of the radiation patterns within jets, neural network-based classifiers are able to out-perform traditional techniques for jet tagging. While these approaches are powerful, neural networks must be applied carefully to avoid performance losses in the presence of domain shift—where the data on which a model is evaluated follows different statistics to the training dataset. This thesis presents studies of possible strategies to mitigate domain shift in the application of deep learning to jet tagging. Firstly, we develop a deep generative model that can separately learn the distribution of quark and gluon jets from mixed samples. Building on the jet topics framework, this model provides the ability to sample quark and gluon jets in high dimension without taking input from Monte Carlo simulations. We demonstrate the advantage of the model over a conventional approach in terms of estimating the performance of a quark/gluon classifier on experimental data. One can also use likelihoods under the model to perform classification that is robust to outliers. We go on to evaluate fully- and weakly-supervised classifiers using real datasets collected at the CMS experiment. Two measurements of the quark/gluon mixture proportions of the datasets are made under different assumptions. Compared to the predictions based on simulation, we either over- or under-estimate the quark fractions of each sample depending on which assumption is made. When estimating the discrimination power of the classifiers in real data we find that while the absolute performance depends on the choice of fractions, the rankings among the models are stable. In particular, weakly-supervised models trained on real jets outperform both simulation-trained models. Our generative networks yield competitive classification and provide a better model for the quark and gluon jet topic distributions in data than the simulation. Finally, we investigate the performance of a number of methods for training mass-generalised jet taggers, with a focus on algorithms that leverage meta-learning. We study the discrimination of jets from boosted Z' bosons against a QCD background and evaluate the networks' performance at masses distant from those used in training. We find that a simple data augmentation strategy that standardises the angular scale of jets with different masses is sufficient to produce strong generalisation. The meta-learning algorithms provide only a small improvement in generalisation when combined with this augmentation.
-
ItemSearching for Dark Matter( 2022)The nature of Dark Matter (DM) is one of the most prominent unanswered questions in particle physics. The Standard Model (SM) has been remarkably successful in describing subatomic phenomena, however not all observations can be explained using this model including DM. The existence of DM is supported by a number of independent astrophysical observations, which when taken together, indicate DM is an elementary particle or particles, however their nature remains largely unknown. The focus of this thesis is on work towards experimental searches for particle DM under the Weakly Interacting Massive Particle (WIMP) paradigm using alter- native but complementary methods to the astrophysical observations in order to test the particle nature of DM. The first approach used is collider searches which test for DM production from the incident SM particles in particle colliders. The second approach is Direct Detection (DD) which aims to observe DM scattering off a SM particle. Using the motion of the Earth and Sun, some more unique features of the expected DM signal may be used to enhance experimental sensitivity. The orbit of the Earth around the Sun results in a time dependent signal with period of a year. The large dataset collected by A Toroidal LHC ApparatuS (ATLAS) allows searches for DM in many areas of phase space. These searches are limited by the ability to identify and discriminate the hypothesised DM signals from back- ground. As such the reconstruction and proper identification of objects in the detector over the largest possible range of momenta plays a key role in what is experimentally accessible. The use of track-jets to allow the identification of low momentum b-hadrons as well as the extension of this identification to lower momentum ranges will be described. The pioneering use of these detector objects to search for DM in regions of phase space previously thought to be inaccessible or too difficult will be described. Many experiments have failed to find DM using direct detection but only one (DArk MAtter (DAMA)) still maintains they have found it, appearing as a time dependent signal. This result is somewhat at odds with other results, however due to experimental differences, it is not completely incompatible. To properly test this an independent experiment using the same experimental approach as DAMA is needed to verify the results. This is the aim of the Sodium-iodide with Active Background REjection (SABRE) experiment, the creation of data acquisition and management systems will be described as well as simulation results used to inform the design and understand detector backgrounds.