Hunting Resonances: Deep Learning for Dijet Searches at CERN's CMS

In high-energy collider physics, identifying new heavy resonances decay is similar to searching for a whisper in a hurricane. At the Large Hadron Collider (LHC) at CERN, protons collide at center-of-mass energies of 13.6 TeV, generating a colossal stream of hadrons. These collimated sprays of particles, known as jets, are the signature of quarks and gluons.

The Challenge of Jet Pairing

In searches for dijet resonances (where a new hypothetical particle decays into a pair of jets), the presence of multiple jets in the final state leads to combinatoric ambiguities. If an event has four jets, which pair belongs to the resonance? Traditional methods rely on simple geometric cuts, but they suffer from high background contamination.

Enter Multi-Layer Perceptrons

During my MSc thesis work, in cooperation with the CMS Experiment, we leveraged deep learning classifiers to optimize this selection. By training Multi-Layer Perceptrons (MLPs) on kinematic features (e.g., angular correlations, transverse momentum, invariant masses), we taught the model to score candidate jet pairs. This resulted in a substantial increase in selection efficiency, boosting the experimental sensitivity to paired dijet resonances by up to 25%.

This research highlights the transformative role of machine learning in particle physics, helping us push the boundaries of the Standard Model.