Pileup Resistant MET using BDT Regression in the ATLAS Level-1 Topological Trigger

Currently, the LHC collides 40 million bunches of protons per second, which produces a petabyte of data every second. To isolate important data, the ATLAS experiment uses a multi-level trigger system: first, the level-1 (L1) trigger uses custom electronics to greatly reduce the rate of incoming data from 40 MHz to 100 kHz; afterwards, the high-level trigger (HLT) uses software based analysis algorithms on PCs to filter out more complex background events, further reducing the rate to roughly 1 kHz. We have developed a package (fwXmachina) that enables the implementation of deep boosted decision trees in the L1 trigger. This allows machine learning algorithms involving regression and classification to run in custom electronics at the nanosecond scale. Using fwX, we present a regression algorithm that can be implemented in the Level-1 Topological (L1Topo) trigger which takes inputs from the various L1Topo missing transverse energy (MET) algorithms and produces a more pileup resistant estimate for MET than any individual algorithm. Pileup makes it more difficult to reconstruct events; therefore, pileup resistant algorithms will continue to be useful in the future as the LHC is upgraded.