Skip to main content

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


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.

This is a preview of subscription content, access via your institution.

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10


  1. A jet is a spray of hadrons, typically originating from the hadronization of gluons and quarks produced in the proton collisions.

  2. In this paper, we set units in such a way that c = \(\hbar\) = 1.


  1. Aaboud M et al (2017) Performance of the ATLAS trigger system in 2015. Eur Phys J C 77(5):317

    Article  ADS  Google Scholar 

  2. Adam W et al (2006) The CMS high level trigger. Eur Phys J C 46:605–667

    Article  Google Scholar 

  3. LeCun Y et al (1990) Handwritten digit recognition with a back-propagation network. In: Touretzky DS (ed) Advances in neural information processing systems 2. Morgan-Kaufmann, Burlington, pp 396–404

    Google Scholar 

  4. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780

    Article  Google Scholar 

  5. Cho K et al (2014) On the properties of neural machine translation: encoder-decoder approaches. In: Proceedings of the 8th workshop on syntax, semantics and structure in statistical translation (SSST-8)

  6. Sjöstrand T et al (2015) An introduction to PYTHIA 8.2. Comput Phys Co 191:159–177

    Article  ADS  Google Scholar 

  7. de Favereau J et al (2014) DELPHES 3, a modular framework for fast simulation of a generic collider experiment. JHEP 02:057

    Article  Google Scholar 

  8. Contardo D et al (2015) Technical proposal for the Phase-II upgrade of the CMS detector. CERN-LHCC-2015-010. LHCC-P-008. CMS-TDR-15-02

  9. Cacciari M, Salam GP, Soyez G (2012) FastJet user manual. Eur Phys. J C72:1896

    Article  ADS  Google Scholar 

  10. Cacciari M, Salam GP, Soyez G (2008) The anti-\(k_t\) jet clustering algorithm. JHEP 04:063

    Article  ADS  Google Scholar 

  11. Madrazo CF et al (2017) Application of a convolutional neural network for image classification to the analysis of collisions in High Energy Physics. Preprint at

  12. Paszke A et al (2017) Automatic differentiation in PyTorch. NIPS Autodiff Workshop. Preprint at

  13. Chollet F et al (2015) Keras. GitHub.

  14. Al-Rfou R et al (2016) Theano: a Python framework for fast computation of mathematical expressions. Preprint at

  15. Kingma DP, Adam JB (2014) A method for stochastic optimization. In: Proceedings of the 3rd international conference on learning representations (ICLR 2015)

  16. Anderson D, Spiropulu M, Vlimant JR (2017) An MPI-based Python framework for distributed training with Keras. Preprint at

  17. Pedregosa F et al (2011) Scikit-learn: machine learning in Python. J Mach Learn Res 12:2825–2830

    MathSciNet  MATH  Google Scholar 

  18. Nair V, Hinton GE (2010) Rectified linear units improve restricted Boltzmann machines. Proc ICML 27:807–814

    Google Scholar 

  19. Huang G et al (2017) Densely connected convolutional networks. In: Proceedings of the 2017 IEEE conference on computer vision and pattern recognition (CVPR). IEEE, Honolulu, HI, pp 2261–2269

  20. Catani S et al (1993) Longitudinally invariant \(K_t\) clustering algorithms for hadron hadron collisions. Nucl Phys B 406:187–224

    Article  ADS  Google Scholar 

  21. Baldi P, Sadowski P, Whiteson D (2014) Searching for exotic particles in high-energy physics with deep learning. Nat Commun 5(07):4308

    Article  ADS  Google Scholar 

  22. Radovic A et al (2018) Machine learning at the energy and intensity frontiers of particle physics. Nature 560(7716):41–48

    Article  ADS  Google Scholar 

  23. de Oliveira L et al (2016) Jet-images—deep learning edition. JHEP 07:069

    Article  Google Scholar 

  24. Guest D et al (2016) Jet flavor classification in high-energy physics with deep neural networks. Phys Rev D 94(11):112002

    Article  ADS  Google Scholar 

  25. Macaluso S, Shih D (2018) Pulling out all the tops with computer vision and deep learning. JHEP 10:121

    Article  ADS  Google Scholar 

  26. Datta K, Larkoski AJ (2018) Novel jet observables from machine learning. JHEP 03:086

    Article  ADS  Google Scholar 

  27. Butter A et al (2018) Deep-learned top tagging with a Lorentz layer. Sci Post Phys 5(3):028

    Article  ADS  Google Scholar 

  28. Kasieczka G et al (2017) Deep-learning top taggers or the end of QCD? JHEP 05:006

    Article  ADS  Google Scholar 

  29. Komiske PT, Metodiev EM, Schwartz MD (2017) Deep learning in color: towards automated quark/gluon jet discrimination. JHEP 01:110

    Article  ADS  Google Scholar 

  30. Schwartzman A et al (2016) Image processing, computer vision, and deep learning: new approaches to the analysis and physics interpretation of LHC events. J Phys Conf Ser 762(1):012035

    Article  Google Scholar 

  31. Bhimji W et al (2018) Deep neural networks for physics analysis on low-level whole-detector data at the LHC. J Phys Conf Ser 1085:042034.

    Article  Google Scholar 

  32. Louppe G et al (2018) QCD-aware recursive neural networks for jet physics. J Phys Conf Ser 1085:042034

    Article  Google Scholar 

  33. Egan S et al (2019) Long short-term memory (LSTM) networks with jet constituents for boosted top tagging at the LHC. JHEP 1901:057

    Google Scholar 

  34. Cheng T (2018) Recursive neural networks in quark/gluon tagging. Comput Softw Big Sci 2(1):3

    Article  Google Scholar 

  35. Gligorov VV, Williams M (2013) Efficient, reliable and fast high-level triggering using a bonsai boosted decision tree. J Instrum 8(02):P02013

    Article  Google Scholar 

  36. Likhomanenko T et al (2015) LHCb topological trigger reoptimization. J Phys Conf Ser 664(8):082025

    Article  Google Scholar 

  37. Beauchemin P-H (2018) Real time data analysis with the ATLAS Trigger at the LHC in Run-2. In: Proceedings of 21st IEEE real time conference (RT2018)

  38. Acosta DE et al (2017) Boosted decision trees in the level-1 muon endcap trigger at CMS. J Phys Conf Ser 1085:042042

    Article  Google Scholar 

  39. Lin J et al (2018) Boosting \({H} \rightarrow b\bar{b}\) with machine learning. JHEP 10:101

    Article  ADS  Google Scholar 

  40. Kuznetsov V (2018) Tensorflow as a service (TFaaS). GitHub.

Download references


This work is supported by Grants from the Swiss National Supercomputing Center (CSCS) under project ID d59, the United States Department of Energy, Office of High Energy Physics Research under Caltech Contract No. DE-SC0011925, and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (Grant agreement no. 772369). T.N. would like to thank Duc Le for valuable discussions during the earlier stage of this project. We thank CERN OpenLab for supporting D.W. during his internship at CERN. We are grateful to Caltech and the Kavli Foundation for their support of undergraduate student research in cross-cutting areas of machine learning and domain sciences. Part of this work was conducted at “iBanks”, the AI GPU cluster at Caltech. We acknowledge NVIDIA, SuperMicro and the Kavli Foundation for their support of “iBanks”.

Author information

Authors and Affiliations


Corresponding author

Correspondence to T. Q. Nguyen.

Ethics declarations

Conflict of interest

On behalf of all authors, the corresponding author states that there is no conflict of interest.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix A An Alternative Use Case

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
figure 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.

Rights and permissions

Reprints and Permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Nguyen, T.Q., Weitekamp, D., Anderson, D. et al. Topology Classification with Deep Learning to Improve Real-Time Event Selection at the LHC. Comput Softw Big Sci 3, 12 (2019).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: