Jonas Eppelt ; Prof. Dr. Torben Ferber ; Prof. Dr. Günter Quast

2022
Insitut for Experimental Particle Physics Karlsruhe

Abstract: In recent years a complementary search paradigm to classical searches has gained traction in the high energy physics community. While classical searches select regions of interest in the data based on a simulation of potential signal models, anomaly detection seeks to identify data regions, that are anomalous with respect to the data. Such searches offer the benefit of independence from concrete models. This thesis explores three autoencoder architectures as a machine-learning-driven tool for model-independent searches at the Belle II experiment. Autoencoders are types of neural networks that compress information in a lower dimensional latent space and reconstruct the information from it. The three architectures explored in this thesis use unregularised latent spaces, latent spaces with a Gaussian prior and with a Dirichlet prior. With the process of encoding and decoding optimized for certain samples, a higher error in the reconstruction is expected for rare samples or samples not represented during the training. This thesis studies the sensitivities of this error for two free parameters of the inelastic Dark Matter model with Dark Higgs. In a comparison between the three architectures unregularized latent spaces provide the highest sensitivities. Studies on the dimensionality of the latent spaces show varying sensitivities for different mass configurations. The studies yield the highest sensitivities for small-mass configurations with an 8-dimensional latent space and for large-mass configurations with a 9-dimensional latent space. Further studies on using the latent space to identify anomalies and efforts to validate the autoencoders on Belle II data are presented.

Note: Presented on 14 11 2022
Note: MSc