Machine Learning and Optimization of LHC Real- Time Event Stream Filter for New Physics Discoveries

Tatiana Likhomanenko (speaker)
Russia, National Research University Higher School of Economics, Yandex Data Factory
Dr. Andrey Ustyuzhanin
Yandex, Yandex School of Data Analysis
LHCb is one of four experiments at LHC in which b-physics is studied. Because the raw information volume is huge, like hundreds GB per second, and it is impossible to store it, an effective filtering system, the so-called trigger system, is necessary to select and store only “interesting” events containing promising decays. That is why a trigger system should reduce the rate at which events are kept for offline processing down to around 2.5-4 kHz and still provide high signal efficiency. Thus, machine learn- ing-powered methods play a crucial role in the trigger system and can improve its efficiency. We upgrade the topological trigger system combining machine learning techniques and physics data structure knowledge for LHC Run 2. The special methods were introduced to improve signal efficiencies for a wide range of b-hadron decays. After that, two different approaches, bonsai BDT and post-pruning, are applied and compared
to speed up online processing in the trigger system. As a result, we get 15%-80% improvement in signal efficiency over the previous Run 1 trigger system depending on the output rate and decay type.