How Machine Learning can exclude your new particle within milliseconds

23 september 2016

In the search for new particles and forces at the Large Hadron Collider (LHC) in Geneva, theoretical models are tested against experimental results in order to identify which models are excluded and which are still going strong. Full models typically have many independent parameters. Unfortunately, such a test can take up to hours for each single model point, making it impossible to sample and analyse a huge high-dimensional model space. Furthermore, such an analysis can be performed exclusively within the LHC experimental collaborations (ATLAS and CMS), since their detector simulation and software framework are not made public. Therefore many physicists currently concentrate their efforts only on simplified low-dimensional models.

Our BSM-AI project aims to solve this issue with the help of Machine Learning (ML). The correct decision (is the model excluded or not) as determined by the experimental collaborations is learned by our ML tool with training data. For the most widely used model for new physics (a high-dimensional version of Supersymmetry called pMSSM) a first tool has been created, called SUSY-AI. Information released by ATLAS was used to train SUSY-AI. Currently SUSY-AI correctly classifies within miliseconds already 93% of randomly chosen models. Moreover, our tool provides a confidence that its classification was correct. The 7% of wrong predictions correlate with a low confidence of SUSY-AI.
Our next steps include a further improvement of the current algorithm by generating more training data and generalising this method to other models (especially models for the elusive dark matter particles).

The research is submitted for publication and can be found on arXiv . SUSY-AI software is made publicly available. SUSY-AI has also an online interface, allowing quick low-volume tests.

SUSY-AI is part of the BSM-AI project which is a collaborative effort of RU Nijmegen and Nikhef (Dr. Sascha Caron and Bob Stienen), the University of Valencia and the University of Madrid. For future developments we cooperate with the Machine Learning group of Tom Heskes (RU Nijmegen).

More information
Dr. Sascha Caron
Email
IMAPP – Particle Physics, RU Nijmegen

Source: Radboud University
SUSY Primer image: copyright Bob Stienen