Roy Cruz
Status: Current Trainee
- GPU training module
- Software Engineering training module
- FPGA training module (in progress)
Biography and Interests
I am a first-year graduate student in the CMS group at the University of Wisconsin-Madison. I am advised by Prof. Tulika Bose.
Project
AXOL1TL is a variational autoencoder-based hardware trigger deployed in Run 3 which identifies anomalous events relative to typical SM events using L1 Global Trigger physics objects. While current High Level Trigger (HLT) integration simply passes the AXOL1TL trigger flag, the project I am contributing to seeks to develop a complementary HLT anomaly detection strategy. The strategy will consist of a transformer encoder-based foundation model trained on a diverse dataset of SM processes to generate low-dimensional embeddings. These representations will be the basis for anomaly scoring or other downstream tasks. Looking forward, the second phase of this project will adapt this approach for the Phase-2 Level-1 Trigger system.
Recent Accomplishments
- Implemented a modified model architecture optimized for reduced resource consumption. Tests confirm the model meets strict HLT latency constraints with negligible performance costs.
- Validation studies show the model’s anomaly detection performance, demonstrating an increase in S/B for various rare SM and BSM processes.
Mentors
- Tulika Bose (University of Wisconsin-Madison)
- Abhijith Gandrakota (Fermilab)