Advancing Machine Learning for High Energy Physics: From Jet Calibration Optimisation to Time Series Anomaly Detection

The ATLAS experiment is one of the two multi-purpose particle collision detectors at the Large Hadron Collider (LHC) at CERN. It generates unprecedented amounts of data, recording proton-proton and nuclei collisions at rates up to 40 MHz. To properly exploit the data and measure the parameters of the Standard Model of particle physics, or even search for new phenomena, we need efficient and precise algorithms. This thesis investigates machine learning-based methods to address two central challenges in ATLAS data processing: jet energy calibration and data quality monitoring of the calorimeter system. These are important issues as they constitute some of the first bottlenecks to collecting high-quality data in hadron collider experiments.

First, we propose a unified approach to jet calibration by merging several steps of the current ATLAS calibration chain. We combine a neural network-based calibration with a loss function that optimises the jet energy resolution (JER) directly in data, using the dijet asymmetry as a proxy. This approach aims to reduce the resources required for deriving the final jet calibration while maintaining or improving the JER performance.

Second, we explore anomaly detection in multivariate time series for data quality monitoring. Accurately identifying the instances when defects occur is essential but challenging, as the types of anomalies are unknown beforehand and reliably labelled data are scarce. We evaluate unsupervised transformer-based models and benchmark their performance against traditional methods on public and ATLAS liquid argon calorimeter data. To address the lack of reliable labels, we use the Lorenzetti simulator – a general-purpose framework for simulating high-energy calorimeters – where we introduce artificial defects to evaluate the sensitivity of various detection methods. As part of the SMARTHEP European Training Network, this project was done in collaboration with IBM Research, who share an interest in time series anomaly detection due to its potential applications in fraud detection.

These contributions highlight how modern machine learning techniques can enhance detector performance and data reliability in particle physics experiments while demonstrating broader interdisciplinary potential.