Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Unsupervised anomaly detection for the ATLAS Level-1 Trigger
The Standard Model has governed particle physics for over four decades, yet certain physical phenomena remain unexplained. Therefore, the search for new processes that could elucidate gaps in our current understanding is crucial. At CERN's Large Hadron Collider, search efforts mainly focus on discov...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English English |
| Published: |
Department of Electrical Engineering
2026
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | The Standard Model has governed particle physics for over four decades, yet certain physical phenomena remain unexplained. Therefore, the search for new processes that could elucidate gaps in our current understanding is crucial. At CERN's Large Hadron Collider, search efforts mainly focus on discovering scientifically well-motivated exper-imental signatures. Yet, in the absence of predefined targets, reliance on models may create blind spots in the data. Searching these blind spots could potentially reveal new physics, a possibility that is especially compelling when considering the low-level data directly read out by the ATLAS detector. The detector collects far more data than can be processed, resulting in over 99% of all data being deleted in real time by the Level-1 Trigger—a chain of field-programmable gate arrays optimized to accept data relevant to the physics processes under study and reject unwanted data. Hundreds of millions of events are rejected every second, possibly discarding something new. Anomaly detection has become a popular approach for searching for new physics without depending on theorized models, thereby maximizing search sensitivity. Deep learning models based on autoencoders have been researched as mechanisms for detecting specific anomalies. However, while these autoencoder-based methods are effective in represen-tation learning and reconstruction, they may fall short in providing a tailored solution for anomaly detection. These methods rely on the availability of clean training data to teach the model what “normal” samples are. This requirement necessitates the develop-ment of large, curated datasets, which would inhibit the development and flexibility of an anomaly detection-based Level-1 Trigger. Furthermore, a preselected background may introduce bias into the detection algorithm, thereby reintroducing model dependence. A Latent Outlier Exposure-based Level-1 Trigger is proposed to train an anomaly detector in the presence of unlabeled physics anomalies. Latent Outlier Exposure involves simul-taneously inferring a binary label for each data point, indicating whether it is anomalous, while updating the model parameters. This is achieved by applying a combination of two losses that share parameters: one for the inferred normal data and one for the inferred anomalous data. This approach was tested on three different anomaly detection systems, including a novel modification to the variational autoencoder's reparameterization trick tailored for anomaly detection. The models were tested on a dataset containing a mix-ture of simulated Standard Model particle content and postulated, but still unobserved, particle content. Experimental results reveal substantial benefits, especially in addressing the formidable challenge of developing an effective, signal-agnostic Level-1 Trigger. |
|---|