Performance of the End-to-End Neural Network Approach to Phase-2 Level-1 Trigger Primary Vertex Reconstruction and Track to Vertex Association

The Phase-2 upgrade of the Level-1 (L1) Trigger will see tracker tracks reconstructed and passed down to global triggering algorithms such as primary vertex (PV) finding. This will be used to associate tracks and other trigger objects to the PV in an event, reducing the impact of pile-up (PU) and increasing trigger performance in certain scenarios. The baseline approach uses a $ \mathrm{ p_T } $-weighted histogram of tracks in $ \mathrm{z_0} $ followed by a 3 bin convolution to find the peak. A set of $ \mathrm{ \eta } $-based cuts are used to associate tracks. The end-to-end neural network approach trains one network on track features to regress the PV position and the classification of PV tracks. The results presented here show relevant features of the L1 tracks and the relationships learnt by the neural network as well as the comparison of the end-to-end approach to the baseline in key metrics. The plots shown are performed with 200K t$ \mathrm{\Bar{t}} $ to semi-leptonic events.