Yapay öğrenme ile jet enerji düzeltmeleri

The LHC at CERN accelerates and collides protons at enormous energies and currently is the most powerful particle accelerator available. The CMS detector is a sophisticated device designed to record the cascade of particles arising from these collisions. With the aid of Monte Carlo simulations, detected particles are reconstructed and clustered together to create physics objects known as jets. The importance of jets is linked to the indirect study of quarks and gluons, since they are not observed freely in nature. Tasks such as reconstruction of high-energy processes, identification of particle interactions and search for new physics beyond the Standard Model all rely on precise study of jets. Due to the detector response and additional factors, the measured energies of the jets need to be calibrated according to simulated truth values. Hence, in the CMS experiment, a sequence of jet energy correction methods is applied to align the jet energies with the true values. In this thesis, we utilized a deep learning approach to further improve standard jet corrections. In recent years, the incorporation of machine learning algorithms has shown great results in high-energy physics studies. Tasks such as jet tagging and particle reconstruction have benefited from recent developments in deep learning. For the jet energy correction task, which is a regression problem, we used two models: the Deep Sets-based Particle Flow Network and a more straightforward fully connected deep neural network. The results of the corrections are presented primarily by two metrics that allow for a comprehensive understanding of performance: average response and relative jet energy resolution. The models surpassed standard energy corrections on a large scale in both metrics.