[Paper Review] CWoLa Hunting: Extending the Bump Hunt with Machine Learning

Jack H. Collins, Kiel Howe+1 more|arXiv (Cornell University)|May 7, 2018

Particle physics theoretical and experimental studies5 references8 citations

TL;DR

This paper introduces a model-independent machine learning method, CWoLa Hunting, that extends traditional bump hunts by leveraging all available multivariate information without prior signal assumptions. By combining classifier-based signal enhancement with invariant mass reconstruction, it achieves superior sensitivity in detecting new resonances—demonstrated on a challenging all-hadronic channel where conventional techniques fail.

ABSTRACT

The oldest and most robust technique to search for new particles is to look for `bumps' in invariant mass spectra over smoothly falling backgrounds. This is a powerful technique, but only uses one-dimensional information. One can restrict the phase space to enhance a potential signal, but such tagging techniques require a signal hypothesis and training a classifier in simulation and applying it on data. We present a new method for using all of the available information (with machine learning) without using any prior knowledge about potential signals. Given the lack of new physics signals at the Large Hadron Collider (LHC), such model independent approaches are critical for ensuring full coverage to fully exploit the rich datasets from the LHC experiments. In addition to illustrating how the new method works in simple test cases, we demonstrate the power of the extended bump hunt on a realistic all-hadronic resonance search in a channel that would not be covered with existing techniques.

Motivation & Objective

• To develop a model-independent method for new physics searches that fully utilizes multivariate data without prior signal hypotheses.
• To overcome the limitations of traditional bump hunts, which rely only on one-dimensional invariant mass distributions.
• To enable detection of resonances in challenging final states—like all-hadronic decays—where standard tagging techniques are ineffective.
• To demonstrate the method’s effectiveness in realistic LHC search scenarios with high background complexity.

Proposed method

• The method uses a classifier trained on simulated signal and background events to identify regions of phase space with enhanced signal sensitivity.
• It applies the classifier output to data, transforming multivariate event information into a one-dimensional discriminant for bump hunting.
• The discriminant is used to construct an invariant mass distribution, enabling standard bump-hunting techniques on a signal-enhanced observable.
• The approach is designed to be sensitive to any new resonance, regardless of its quantum numbers or decay topology, ensuring full coverage.

Experimental results

Research questions

• RQ1 Can a machine learning approach extend traditional bump hunts to use all available multivariate information without assuming a signal model?
• RQ2 How does the method perform in detecting resonances in final states where standard tagging techniques fail, such as all-hadronic decays?
• RQ3 To what extent does the method improve sensitivity compared to conventional bump hunts in realistic LHC search environments?
• RQ4 Can the method detect new physics signals in high-background, complex phase spaces without prior signal hypotheses?

Key findings

• The method successfully detects resonances in an all-hadronic final state where conventional bump hunts and tagging techniques fail due to overwhelming QCD backgrounds.
• By using multivariate information through a trained classifier, the method enhances signal sensitivity beyond what is possible with invariant mass alone.
• The approach achieves full coverage of the phase space without requiring assumptions about signal topology or quantum numbers.
• The technique demonstrates that model-independent searches can be significantly more powerful than traditional one-dimensional bump hunts in complex, realistic LHC environments.