cern.ch

Real-time discrimination of photon pairs using machine learning at the LHC

[to restricted-access page]

Information

Tools

Abstract

ALP-mediated decays and other as-yet unobserved $B$ decays to di-photon final states are a challenge to select in hadron collider environments due to the large backgrounds that come directly from the $pp$ collision. We present the strategy implemented by the LHCb experiment in 2018 to efficiently select such photon pairs. A fast neural network topology, implemented in the LHCb real-time selection framework achieves high efficiency across a mass range of $4-20$ GeV$/c^{2}$. We discuss implications and future prospects for the LHCb experiment.

Figures and captions

Output probability of the classifier for the different topologies. Upper left: 0CV, upper right: 1CV DD, lower 1CV LL

three_[..].pdf [20 KiB]
HiDef png [246 KiB]
Thumbnail [195 KiB]
three_plots.pdf

Signal and background distributions of information used to train the 1CV LL classifier. Variables are explained in the text.

LL_convm.png [5 KiB]
HiDef png [68 KiB]
Thumbnail [25 KiB]
LL_convm.png
LL_xpt.png [5 KiB]
HiDef png [71 KiB]
Thumbnail [26 KiB]
LL_xpt.png
LL_xipchi2.png [5 KiB]
HiDef png [64 KiB]
Thumbnail [25 KiB]
LL_xipchi2.png
LL_gprob.png [6 KiB]
HiDef png [67 KiB]
Thumbnail [24 KiB]
LL_gprob.png
LL_ge49.png [6 KiB]
HiDef png [75 KiB]
Thumbnail [27 KiB]
LL_ge49.png
LL_ptasym.png [5 KiB]
HiDef png [69 KiB]
Thumbnail [23 KiB]
LL_ptasym.png

Signal and background distributions of information used to train the 1CV DD classifier. Variables are explained in the text.

DD_convm.png [6 KiB]
HiDef png [78 KiB]
Thumbnail [28 KiB]
DD_convm.png
DD_xpt.png [6 KiB]
HiDef png [72 KiB]
Thumbnail [27 KiB]
DD_xpt.png
DD_gprob.png [5 KiB]
HiDef png [70 KiB]
Thumbnail [25 KiB]
DD_gprob.png
DD_ge49.png [5 KiB]
HiDef png [71 KiB]
Thumbnail [25 KiB]
DD_ge49.png
DD_ptasym.png [5 KiB]
HiDef png [68 KiB]
Thumbnail [24 KiB]
DD_ptasym.png

Signal and background distributions of information used to train the 0CV classifier. Variables are explained in the text.

None_xpt.png [6 KiB]
HiDef png [79 KiB]
Thumbnail [31 KiB]
None_xpt.png
None_m[..].png [5 KiB]
HiDef png [70 KiB]
Thumbnail [25 KiB]
None_mingprob.png
None_m[..].png [6 KiB]
HiDef png [72 KiB]
Thumbnail [27 KiB]
None_maxgprob.png
None_m[..].png [6 KiB]
HiDef png [77 KiB]
Thumbnail [29 KiB]
None_minShowerShape.png
None_m[..].png [6 KiB]
HiDef png [81 KiB]
Thumbnail [31 KiB]
None_maxShowerShape.png

ROC curves for the test data using the different topologies. 0CV NN (left) , 1CV DD NN (center),1CV LL NN (right) .

ROC_curves.pdf [79 KiB]
HiDef png [78 KiB]
Thumbnail [48 KiB]
ROC_curves.pdf

Animated gif made out of all figures.

DP-2019-004.gif
Thumbnail
thumbnail_DP-2019-004.gif

Tables and captions

Selection applied in the \texttt{Hlt1B2GammaGamma} and \texttt{Hlt1B2GammaGammaHighMass} HLT1 trigger selection. Energies and masses given here are computed with $2\times2$ cell clusters.

Table_1.pdf [53 KiB]
HiDef png [64 KiB]
Thumbnail [32 KiB]
tex code
Table_1.pdf

Percentage efficiency relative to all candidates accepted by the Photon and Electron channels of the L0 hardware trigger for the $ B ^0_ s $ and ALP samples, combining all the $\gamma$ reconstruction modes.

Table_2.pdf [51 KiB]
HiDef png [80 KiB]
Thumbnail [40 KiB]
tex code
Table_2.pdf

Sample sizes for the signal decays and background after reconstruction and trigger requirements.

Table_3.pdf [49 KiB]
HiDef png [64 KiB]
Thumbnail [29 KiB]
tex code
Table_3.pdf

Kolmogorov-Smirnov p-values for the comparison of the classifier's distribution between the test and training samples for the different topologies

Table_4.pdf [29 KiB]
HiDef png [49 KiB]
Thumbnail [24 KiB]
tex code
Table_4.pdf

Percentage efficiency for the $ B ^0_ s $ and ALP samples relative to the reconstructed and loosely selected samples.

Table_5.pdf [54 KiB]
HiDef png [84 KiB]
Thumbnail [41 KiB]
tex code
Table_5.pdf

Created on 12 October 2019.