Developing a Hybrid Machine Learning Model for VELO Upgrade Track Reconstruction

High energy physics experiments have been processing large quantities of data for decades. The growth of high volume data collection in industry has led to significant technical development in big data and machine learning due to their enormous commercial potential. This knowledge can be harnessed by the next generation of high energy physics experiments, which require sophisticated real-time, high throughput data processing pipelines to exploit the physics potential of huge datasets. The ambitious goals of the upgrade LHCb experiment require a shift to reconstructing all particle collisions in real-time. The corresponding increase in data rate will test the capability of conventional processing algorithms, such as the methods of track reconstruction. Machine learning methods are potential candidates for future tracking algorithms. They can learn from data without prior physics knowledge and are capable of incredible speed through parallelisation on specialist hardware such as FPGAs. In this thesis I focus on the VELO detector, which surrounds the proton collision point at the LHCb experiment. Tracks reconstructed in the VELO are a vital component in reducing data throughput by the rejection of unwanted collision events. After experimenting with machine learning methods and toy tracking models, I built on work to develop a hybrid VELO tracking algorithm incorporating a neural network to join pairs of hits into track seeds. The algorithm matches baseline performance target, and is more robust to small random detector misalignment than the conventional tracking algorithm.