Machine learning for top quark physics

In the last 5 years, machine learning algorithms, in particular the neural network, have proven to be a very powerful tool for high energy physics at the LHC. In the realm of top quark physics, machine learning has risen to prominence both in event selection and event reconstruction.

In this thesis a Deep Neural Network for jet flavour identification is presented. It is capable of identifying b jets, c jets, light quark jets and gluon jets. By utilizing a novel neural network architecture that can efficiently exploit the full jet information it achieves state of the art performance in each of the jet classification tasks. This neural network is extended further to estimate jet energy corrections. Since the jet energy response depends on the flavour of the jet, the architecture and inputs for jet flavour identification can be utilized for making a general jet energy correction, which leverages the flavour information going beyond the standard approach of jet energy corrections that has no jet flavour dependence.

Machine learning has not just found success for particle physics re-construction, but it is also heavily used for particle physics event selection. In this case the machine learning methods are optimized to minimize the statistical uncertainty on the measurement by increasing selection efficiency and reducing the rate of background. However in modern particle physics analyses the systematic uncertainty is the dominant component of the total uncertainty. In this thesis a novel machine learning method is developed that makes an event selection that reduces the systematic uncertainty. A showcase of the method is done in the setting of a top quark mass measurement.