Knowledge Distillation of a Normalising Flow for Real-Time Anomaly Detection at LHC Level-1 Trigger
Abstract
We present a new strategy for unsupervised anomaly detection at the hardware-based first stage of event processing (Level-1 trigger) of the Large Hadron Collider (LHC). A normalising flow trained exclusively on Standard Model (SM) events provides a teacher anomaly score based on its exact negative log-likelihood, which is distilled into a compact neural network suitable for field-programmable gate array (FPGA) deployment. The teacher reaches state-of-the-art performance on a benchmark dataset, using a rigorous p-value-based definition of the anomaly score. A simple two-hidden-layer student reproduces the teacher performance across four beyond-the-SM benchmarks to within 0.1 percentage points in area under the receiver operating characteristic curve (AUC), while achieving a compression factor of approximately 325 times. Quantisation-aware training with PQuantML reduces the model to 8-bit precision with AUC changes below 0.06 percentage points. Compilation to Verilog firmware with Alkaid yields FPGA designs that require neither digital signal processors nor block random-access memory and achieve latencies of 27-52 ns. This modular pipeline decouples the expressiveness of the anomaly-detection model from real-time hardware constraints, enabling complex architectures to be used for model-agnostic searches for new physics at the full LHC collision rate of 40 MHz.
Related
- A temperature-adaptive logit-standardized knowledge distillation for lightweight network intrusion detection
- Information-Theoretic Reliability--Safety Bounds and Finite-Alignment Dynamics in AI Knowledge Distillation
- Anomaly detection for multijet scenarios
- Domain Adaptation against Background Sculpting in Anomaly Detection at the LHC
- Model-independent searches and anomaly detection at the CMS experiment
BibTeX
@misc{abson2026knowledge,
title = {{Knowledge Distillation of a Normalising Flow for Real-Time Anomaly Detection at LHC Level-1 Trigger}},
author = {Jaiman Abson and Florencia Canelli and Valentina Guglielmi and Roope Oskari Niemi and Maurizio Pierini and Chang Sun and Francesco Vaselli},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.15295},
eprint = {2609.15295},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.15295},
note = {CC BY-SA 4.0},
}
RIS
TY - UNPB
TI - Knowledge Distillation of a Normalising Flow for Real-Time Anomaly Detection at LHC Level-1 Trigger
AU - Jaiman Abson
AU - Florencia Canelli
AU - Valentina Guglielmi
AU - Roope Oskari Niemi
AU - Maurizio Pierini
AU - Chang Sun
AU - Francesco Vaselli
PY - 2026
DA - 2026/09/17
PB - arXiv
DO - 10.48550/arxiv.2609.15295
UR - https://arxiv.org/abs/2609.15295
AB - We present a new strategy for unsupervised anomaly detection at the hardware-based first stage of event processing (Level-1 trigger) of the Large Hadron Collider (LHC). A normalising flow trained exclusively on Standard Model (SM) events provides a teacher anomaly score based on its exact negative log-likelihood, which is distilled into a compact neural network suitable for field-programmable gate array (FPGA) deployment. The teacher reaches state-of-the-art performance on a benchmark dataset, using a rigorous p-value-based definition of the anomaly score. A simple two-hidden-layer student reproduces the teacher performance across four beyond-the-SM benchmarks to within 0.1 percentage points in area under the receiver operating characteristic curve (AUC), while achieving a compression factor of approximately 325 times. Quantisation-aware training with PQuantML reduces the model to 8-bit precision with AUC changes below 0.06 percentage points. Compilation to Verilog firmware with Alkaid yields FPGA designs that require neither digital signal processors nor block random-access memory and achieve latencies of 27-52 ns. This modular pipeline decouples the expressiveness of the anomaly-detection model from real-time hardware constraints, enabling complex architectures to be used for model-agnostic searches for new physics at the full LHC collision rate of 40 MHz.
ER -