Unsupervised Machine Learning for Rare Signal Detection in CMS Detector Data

We present here a method for identifying high-p_t beyond the Standard Model (BSM) signals, with no dependence on the expected physics signatures of these models. We do this by constraining our detector data to the expected signal region of a given model and training a deep neural autoencoder to 'recognize' jets belonging to the Standard Model (SM) background data. In this way, our models learn the background signature of SM jets in our signal region. We then evaluate this model on a mixture of SM and BSM jets, flagging jets with a high autoencoder reconstruction error as 'anomalous' or signal jets. This method of searching independently of any physics model is especially useful when the BSM model in question involves particles of unknown mass, branching fraction, etc. This is especially relevant in Dark Matter searches, such as the search for Semi-Visible jets in missing-E_T events at the LHC. In these cases, our models are capable of flagging a range of BSM models with different parameters within our signal region, as opposed to a supervised model (such as a Boosted Decision Tree) which is only able to flag the exact signal it was trained on. We show that this technique is a promising method of identifying arbitrary BSM signals to high precision in LHC data.