Search for new physics signatures in ATLAS data using autoencoders

Searches for new physics beyond the Standard Model are mostly driven by theories predicting new particles or processes. Due to the theoretical bias of these searches, the investigated phase space is very narrow. Thus, a new signal might not be found even if it occupies a phase space close to the investigated one. Autoencoders offer a semi supervised and bias-free method to search for new signals in any region. They are a special type of neural networks, that reproduces the input using a lower dimensional representation. Autoencoders are trained in a semi supervised manner, which allows them to detect anomalies while only being trained on known processes of the Standard Model. This master thesis evaluates autoencoders as a method to search for new physics signatures, using data from $pp$ collisions at $\sqrt{s}=13$ TeV measured by the ATLAS detector. To evaluate autoencoders, a region that is dominated by tt events, with a minor fraction of single top-quark $tW$ events is used. For that region a discriminant called mini-max $m(lb)$ already exists. A detailed comparison between the two algorithms is carried out. The autoencoder learns a partial representation of mini-max $m(lb)$ and additionally selects a region with increased discrepancy, that is not included by mini-max $m(lb)$. Furthermore, an autoencoder is applied to a region with a discrepancy between data and simulation, where the origin of the discrepancy is unknown. The autoencoder suggests that the discrepancy originates from a shape mismatch and/or a scaling error.