Topology Classification with Deep Learning to Improve Real-Time Event Selection at the LHC

Abstract

We show how an event topology classification based on deep learning could be used to improve the purity of data samples selected in real time at the Large Hadron Collider. We consider different data representations, on which different kinds of multi-class classifiers are trained. Both raw data and high-level features are utilized. In the considered examples, a filter based on the classifier’s score can be trained to retain $\sim 99\%$ of the interesting events and reduce the false-positive rate by more than one order of magnitude. By operating such a filter as part of the online event selection infrastructure of the LHC experiments, one could benefit from a more flexible and inclusive selection strategy while reducing the amount of downstream resources wasted in processing false positives. The saved resources could translate into a reduction of the detector operation cost or into an effective increase of storage and processing capabilities, which could be reinvested to extend the physics reach of the LHC experiments.

Appendix A An Alternative Use Case

In this paper, we showed how one could use a topology classifier to keep the overall trigger rate under control while operating triggers with otherwise unsustainable loose selections. In this appendix, we discuss how topology classifiers could be used to save resources for a pre-defined baseline trigger selection by rejecting events associated to unwanted topologies. In this case, the main goal is not to reduce the impact of the online selection. Instead, we focus on reducing resource consumption downstream for a given trigger selection.

To this purpose, we consider a copy of the data set described in section “Data Set”, obtained tightening the $p_\mathrm{{T}}$ threshold from 23 to 25 GeV and the isolation requirement from $\mathtt{ISO}< 0.45\hbox { to} \mathtt{ISO} < 0.20$. Doing so, the sample composition changes as follow: 7.5% QCD; 92% W; 0.5% $t \bar{t}$. With such selections, the trigger acceptance rate would decrease from 690 Hz to 390 Hz, closer to what is currently allocated for these triggers in the CMS experiment.

Following the procedure described in sections “Model description” and “Results”, we train the same topology classifiers on this data set. The corresponding ROC curves are presented in Fig. 11 for a $t \bar{t}$ and a W selector.

Fig. 11

ROC curves for the $t\bar{t}$ (left) and W (right) selectors described in the paper, trained on a data set defined by a tighter baseline selection

We then define a set of trigger filters applying a lower threshold to the normalized score of the classifier, choosing the threshold value that corresponds to a certain TPR value. The result is presented in Table 3, in terms of the FPR and the trigger rate.

Table 3 False-positive rate (FPR) and trigger rate (TR) corresponding to different values of the true-positive rate (TPR), for a $t \bar{t}$ (top) and W selector

The trigger baseline selection we use in this study, close to what is used nowadays in CMS for muons, gives an overall trigger rate (i.e., summing electron and muon events) of $\sim$ 390 Hz (i.e., 190 Hz per lepton flavor). If one was willing to take (as an example) half the W events and all the $t \bar{t}$ events, this number could be reduced to $\sim 200\hbox { Hz}$ using the inclusive selectors presented in this study (taking into account the partial overlap between the two triggers). A more classic approach would consist in prescaling the isolated-lepton triggers, i.e., randomly accepting half of the events. The effect on W events would be the same, but one would lose half of the $t \bar{t}$ events while still writing 15 times more QCD than $t \bar{t}$ events. In this respect, the strategy we propose would allow a more flexible and cost-effective strategy.