Abstract:
In high-energy physics, the very large data volume—with a substantial fraction arising from
background noise—makes it difficult to identify the phenomena of interest. To overcome this
complexity, an online event-selection system (the trigger) is employed, which in the ATLAS
experiment at the LHC operates in two sequential stages: the first-level trigger and the High-Level
Trigger (also called the rapid stage). Machine-learning techniques have been adopted to enhance
this selection.
In the context of electron identification, an accurate estimation of the energy deposited in
the calorimeters is crucial for the correct selection of candidates. Measurement errors can stem
from pileup effects, which artificially increase the reconstructed energy, and from longitudinal
and lateral energy losses, which lead to an underestimation of the true energy. Consequently,
inaccuracies in the energy estimate can compromise selection efficiency or raise the false-positive
rate.
This work proposes an energy-calibration methodology for the High-Level Trigger using an
ensemble of gradient-boosted decision trees (GBDTs). This approach is capable of modeling
non-linearities and capturing complex relationships in the data efficiently, thereby improving the
precision of the energy estimate.
The solution was integrated into the detector’s operational software framework and is currently
under evaluation for potential adoption in the coming years. In tests performed with simulated
data, we observed up to a 25% reduction in the dispersion of reconstructed energy, as well as a
20.6% improvement in the mean absolute percentage error in the low-energy range (0–30 GeV).
With validation data, it was possible to lower the selection threshold without compromising the
electron identification efficiency—resulting in a reduced false-positive rate, lower computational
demand, and an overall increase in trigger performance.
