Next Generation Triggers for CMS: Integrating the Kubeflow platform for Level-1 Trigger MLOps applications

The CMS experiment at the Large Hadron Collider relies on a two-tiered trigger system to reduce the 40 MHz collision rate to approximately 1 kHz for permanent storage. Traditional Level-1 Trigger (L1T) decision logic, based on rules and thresholds, risks missing unexpected physics signals. To address this, CMS has begun exploring machine learning (ML) models such as anomaly detection autoencoders at the trigger level. This internship project investigated the feasibility of using Kubeflow, an open-source MLOps platform, to orchestrate the ML lifecycle for CMS triggers. A prototype pipeline was developed on CERN’s Kubernetes infrastructure, including preprocessing of CMS Level-1 data, training of a Variational Autoencoder (VAE), and evaluation using background and signal datasets. The evaluation demonstrated Kubeflow’s strengths in modularity, reproducibility, and scalability, with additional integrations for GitLab CI/CD and heterogeneous hardware scheduling. The findings suggest that Kubeflow is a promising candidate for CMS ML workflows at the High-Luminosity LHC, though future work is required to integrate a model registry, migrate to S3-compatible storage, and validate FPGA toolchain compatibility.