Connecting the dots

To study LHC events it is necessary to reconstruct each particle's trajectory. Track reconstruction is a difficult problem due to its combinatorial nature and many fake tracks are generated during the process, most of which are eliminated only after many iterations, increasing the computational cost of the algorithm. In order to avoid part of this cost the objective of this project is to create high quality track seeds, filtering the fake tracks as soon as possible. The filter is implemented with a convolutional neural network which takes hit shapes as the main input features to determine the compatibility between different hits. Several network topologies are trained and tested to study the tradeoff between computational cost, model complexity and model performance. Furthermore, the model inference is benchmarked with several hardware architectures