Benchmarking Manifold Learning vs. VAEs for the CMS Level-1 Trigger Using the ADC 2021 Dataset

At the Compact Muon Solenoid (CMS) Experiment at the LHC, the Level-1 Trigger utilizes Field Programmable Gate Arrays (FPGAs) to swiftly filter 40 megahertz streams of proton-proton collision events in order to satisfy low latency and computing resource constraints. Traditionally, it implements fundamental physical thresholds to filter particles of certain types. However, this theory-dependent approach could potentially discard data revealing new physical phenomena. To address this, we explore data-driven unsupervised machine learning techniques for theory-independent anomaly detection. We seek to compare such techniques using the public Anomaly Detection Challenge (ADC) 2021 dataset [1] as a benchmark. Specifically, we compare traditional variational autoencoders (VAEs) with alternative manifold learning strategies. Manifold learning seeks to produce and analyze abstract (latent) spaces as a potential way to identify anomalies and explore whether they represent new physical phenomena or potential detector flaws. We reconstruct input spaces from latent spaces and use receiver operating characteristic (ROC) curves to gauge performance across signals of benchmark particle decay scenarios. Although exploratory, the results of this work could potentially motivate future CMS trigger strategies.

[1] https://arxiv.org/pdf/2107.02157.pdf