Laurea Magistrale

Memristor Technology for High-Energy Physics: Potential and Challenges

The rapid advancement of memristor technology is driven by its numerous promising applications across various domains. One of the most remarkable features of memristors is their ability to store information passively, enabling the development of energy-efficient storage devices that consume no power during idle states. This capability is particularly advantageous for the creation of non-von Neumann computing architectures, where memory and information processing are integrated. Such architectures facilitate faster execution of computational tasks, including neural network inference, by eliminating the bottlenecks associated with traditional memory access.

This project is directly related to the ATLAS experiment at the Large Hadron Collider (LHC) and aims to explore the potential of memristor technology for future detector upgrades. Specifically, their integration into the data acquisition (DAQ) systems of ATLAS could be highly beneficial, as its upgrades require the ability to analyze high event frequencies using advanced algorithms. This would maximize efficiency in identifying physics events of interest within the detection system while also improving data processing efficiency and reducing power consumption.

This is a laboratory-based project that involves direct experimental studies on memristors, both as single devices and within matrices, to characterize their behavior under extreme conditions. The influence of temperature variations and external magnetic fields on memristor conductance must be analyzed through direct laboratory measurements of variations in conductance. These controlled tests will allow for the estimation of key parameters affecting device performance, ensuring accurate characterization of their stability and reliability under extreme conditions. This is essential for evaluating their resilience to radiation exposure and their potential implementation in radiation-hardened computing systems, critical aspects for high-energy physics experiments like ATLAS.

Contacts: Valerio Ippolito, Davide Fiacco

Implementation of a Graph Neural Network on an FPGA for High Luminosity upgrade of the ATLAS Detector

High-energy physics (HEP) experiments at the Large Hadron Collider (LHC) generate an immense volume of data, requiring real-time processing capabilities to extract relevant physics signals efficiently. As the LHC enters the High-Luminosity (HL) era, with an expected increase in instantaneous luminosity by a factor of ten, conventional data acquisition and triggering systems face significant computational challenges. Field-Programmable Gate Arrays (FPGAs) offer a crucial solution due to their inherent parallelism, low-latency processing, and energy efficiency, making them essential for the next-generation upgrades of the ATLAS detector.

Graph Neural Networks (GNNs) have emerged as a powerful tool for pattern recognition in HEP due to their ability to model complex spatial relationships in particle interactions. Unlike conventional convolutional neural networks (CNNs), which operate on regular grid-like data, GNNs can process irregular and non-Euclidean structures, making them particularly suitable for analyzing particle trajectories and detector hits. GNNs have demonstrated superior performance in tracking, jet identification, and anomaly detection, crucial tasks for improving event selection in high-energy physics experiments.

The primary goal of this project is to implement a hardware-efficient GNN on an FPGA for real-time data processing in the ATLAS  detector. This involves designing and optimizing a GNN architecture specifically for FPGA deployment, with a focus on efficient matrix multiplication and sparse graph representation. Additionally, a hardware-friendly inference model will be developed to minimize resource utilization while maintaining high accuracy. The project will integrate the FPGA-based GNN within a simulated detector data pipeline to assess its performance in realistic event processing scenarios. Finally, the implementation will be benchmarked against GPU-based models in terms of latency, power consumption, and inference accuracy.

Contacts: Valerio Ippolito, Davide Fiacco, Giuliano Gustavino, Stefano Giagu

Anomaly detection in hadronic final states with ATLAS Run-3 data

This thesis proposes a comprehensive study on the application of anomaly detection techniques to the analysis of the Run-3 data from the Large Hadron Collider (LHC), specifically focusing on resonant final states from the hadronic decay of boosted particles. The goal is to identify potential signs of new physics by analyzing anomalies in the data that do not conform to the Standard Model predictions. This work will leverage advanced statistical and machine learning (ML) methodologies to enhance the sensitivity of high-energy particle experiments to rare or unexpected events. Additionally, the thesis will explore the development of novel ML techniques aimed at identifying anomalous jets originating from new physics particles, which could be key in discovering beyond the Standard Model phenomena.

Contacts: Stefano Giagu, Valerio Ippolito, Giuliano Gustavino

Anomaly Detection Analysis Using Calorimeter Clusters

The search for new physics at the LHC increasingly relies on data-driven techniques to identify unexpected signatures beyond Standard Model predictions. This thesis focuses on an anomaly detection analysis leveraging calorimeter cluster information to identify unconventional jet signatures. The approach employs unsupervised machine learning methods trained on calorimeter observables to distinguish anomalous jets from the dominant QCD background. The next steps involve defining background and signal regions using statistical analysis techniques, performing a fit of the dijet invariant mass distribution of anomalous jets, and interpreting potential excesses in the context of physics beyond the Standard Model.

Contacts: Giuliano Gustavino

Development of an Anomaly Detection Tagger Based on Tracking Variables

The identification of anomalous jets is a crucial challenge in searches for new physics at the LHC, particularly in scenarios involving non-standard signatures. This thesis focuses on the development of a jet tagging algorithm based on anomaly detection techniques, leveraging tracking variables to enhance discrimination power. The approach exploits information from both prompt and displaced tracks to identify deviations from the Standard Model expectations. The study will involve training and evaluating machine learning models, particularly unsupervised learning methods, to detect rare and unexpected features in jet substructure. The tagger’s performance will be assessed in simulated datasets, with potential applications in real-time trigger selection and offline analyses within ATLAS.

Contacts: Giuliano Gustavino

Pheomenological Studies of Dark QCD: Connecting Models and Reinterpreting Current Searches

Dark QCD scenarios introduce a rich phenomenology, often leading to signatures distinct from conventional new physics searches. This thesis aims to bridge the gap between different theoretical models by studying their common phenomenological features and exploring how current experimental results can be reinterpreted in this context. The work will involve simulating various dark QCD scenarios, identifying key observables, and comparing them with existing searches at the LHC. Additionally, a systematic approach to reinterpret experimental exclusions and sensitivities will be developed to assess their impact on a wide range of theoretical frameworks. The study will provide insights into potential refinements of search strategies to better capture the signatures of dark QCD dynamics.

Contacts: Giuliano Gustavino

Fast scintillators for muon trigger and tracking.

Fast scintillators with wavelength-shifting fibers, readout with Silicon Photomultipliers, are a very promising technology for muon trigger and tracking both for  experiments at a Future Circular Collider and for possible future upgrades of the ATLAS  experiment at the Large Hadron Collider (LHC). The thesis will consist of studying the performance (in terms of  efficiency, timing, resolution) of different types of scintillator-based detectors, with laboratory measurements on  prototypes and possibly tests with particle beams. The work will also include studies and development of a full simulation  of the detector setup using state-of-the-art simulation software, comparisons with data, and simulation of their  usage at collider experiments.

Contacts: Stefano Rosati, Massimo Corradi, Fabio Anulli, Cesare Bini

Development and study of the performances of CNN and Graph Neural Networks on hardware accelerators (FPGAs/ACAPs/GPUs) for the High Level trigger of the ATLAS experiment

Development of state-of-the-art DNN/RNN/CNN/GNN neural architectures for the identification of events caused by interesting physics processes in the high-level trigger (software trigger) of the ATLAS experiment. Implementation of the algorithms on hardware coprocessors (FPGA/ACAP/GPU) using next-generation software libraries for optimization and synthesis (Xilinx VITIS AI, Intel OpenVINO) and performance measurement compared to conventional algorithms.

Contacts: Stefano Giagu

DL and RL based real-time monitoring of ATLAS detector anomalies

The ATLAS detector at the Large Hadron Collider generates an immense volume of data, requiring robust real-time monitoring to detect anomalies that could indicate hardware failures, calibration drifts, or unexpected physical phenomena. This thesis explores the application of reinforcement learning techniques for anomaly detection in ATLAS, leveraging deep RL agents to dynamically adapt to evolving detector conditions. Unlike traditional rule-based or supervised learning approaches, RL can autonomously learn optimal strategies for identifying and classifying anomalies by interacting with real-time detector data streams. The proposed methodology involves training an RL-based anomaly detection system using both simulated and historical data, optimizing its reward function for early detection and minimal false positives. The project aims to enhance the efficiency and reliability of ATLAS operations, contributing to the broader effort of automating high-energy physics data analysis with machine learning-driven solutions.

Contacts: Stefano Giagu

AI assisted data-compression strategies to optimise real-time data processing bandwidth and physics potential of the ATLAS experiment at the LHC

The ATLAS experiment at the Large Hadron Collider produces ~ a petabyte of raw data per second, far exceeding storage and bandwidth constraints. Efficient data compression strategies are crucial to optimizing real-time processing while preserving essential physics information. This thesis investigates the application of AI-assisted data compression techniques, leveraging machine learning models to identify and retain high-value physics data while reducing redundancy. The approach explores deep learning-based autoencoders, reinforcement learning, and transformer architectures to dynamically adapt compression rates based on real-time detector conditions and event significance. By integrating AI-driven methods into the ATLAS data acquisition pipeline, this work aims to maximize physics discovery potential while ensuring efficient use of computational and storage resources. The proposed system will be validated using both Monte Carlo simulations and real ATLAS detector data, with the goal of contributing to the broader effort of intelligent, adaptive data handling in high-energy physics.

Contacts: Stefano Giagu

Development and application of a novel generation of Deep-learning models for applications in High Energy Physics

• Intrinsically explainable modular AI based on transformer and foundation models for analysis, interpretation and recasting of HEP analyses at LHC.
• Large-Language Model based representation learning via multimodal Transformers of physical objects in a High Energy Physics Detector and applications to multiboson measurements and searches.

These two projects focused in innovative research, aiming at either developing a new generation of deep learning models that are intrinsically explainable, based on the recent field of the modular AI, and realizing a digital twin of a collider experiment that can be used for different tasks: generative, anomaly detections, reconstruction etc… The digital twin will be based on a large language model in which the language is represented by the particles’ interactions with the experiment’s detectors. 

Contacts: Stefano Giagu

High-precision modelling of irreducible backgrounds in the Higgs to dimuon search with the ATLAS Experiment

The search for the Higgs boson decaying to a pair of muons is a crucial test of the Standard Model, given the rarity of this process and its sensitivity to new physics. A key challenge in this analysis is the precise parameterisation of the dimuon invariant mass background, which must be controlled at the permille level to distinguish the signal effectively. 

This project explores novel machine learning approaches to improve background modelling and enhance the overall analysis reach. It will focus on developing and evaluating machine learning methods such as Gaussian processes for background modelling, investigating the integration of background modelling into the analysis optimisation pipeline, and implementing adversarial neural networks to refine the description of background distributions.

The study will leverage ATLAS data and Monte Carlo simulations to train and validate machine learning models, evaluating their performance in terms of fit quality, uncertainty reduction, and impact on the Higgs to dimuon search. By improving background parameterisation to the required permille-level precision, this work aims to enhance the sensitivity of the search.

Contacts: Alessandra Betti, Giacomo Artoni

Development of Continuous Muon Momentum Calibrations for the ATLAS Experiment

The precise calibration of muon momentum is essential for many physics analyses in ATLAS, particularly those involving high-precision measurements and searches for rare processes. Traditional calibration methods rely on data-driven techniques and simulation-based corrections, but achieving optimal performance across the detector requires a detailed 

understanding of resolution effects and scale biases. This project explores the use of Machine Learning to model detector performance in muon momentum reconstruction, focusing on the prediction of resolution, scale biases, and other systematic effects. Additionally, it will investigate the feasibility of an ML-based calibration pipeline for muons, aiming for a continuous parameterisation of the relevant scale and resolution correction factor. The study will leverage simulated and real ATLAS data to train and validate ML models, assessing their ability to improve momentum reconstruction across different detector regions and kinematic regimes. By refining the precision of muon momentum calibration, this work has the potential to enhance the sensitivity of key physics analyses, including Higgs boson studies and new physics searches.

Contacts: Luca Martinelli, Giacomo Artoni

Search for non-resonant Higgs boson pair production in the bbττ final state with ATLAS Run 3 data 

After the discovery of the Higgs boson by the ATLAS and CMS Collaborations in 2012, both experiments have started a vaste programme of measurements of the Higgs boson properties to test if they all agree with the Standard Model (SM) predictions, as any deviation of the observed values from the expected ones could be a hint of the presence of new physics beyond the SM (BSM). One very interesting property of the Higgs boson to study is its self-coupling, that is one of the parameters determining the shape of the Higgs potential around the minimum, and that is still largely unconstrained by experimental measurements. The Higgs boson self-coupling can be directly measured at the Large Hadron Collider through the study of non-resonant Higgs boson pair (HH) production, a very rare process predicted by the SM that has not yet been observed and whose properties can strongly be affected by possible BSM values of the Higgs boson self-coupling. 

The proposed thesis project will focus on the search for HH production in the bbττ final state, that is a final state with a very good compromise between relatively high branching ratio and relatively low background, resulting in a good signal over background ratio that makes it one of the most sensitive channels for the HH measurements. The thesis project will be carried out using ATLAS Run 3 data and simulations and will have the goal of testing possible improvements of the sensitivity of the HH->bbττ analysis through the inclusion of additional data recorded via new triggers in new analysis categories and the development of new machine learning algorithms for signal and background classification.

Contacts: Alessandra Betti

Development of new machine-learning based methods for di-τ mass reconstruction in ATLAS

Many interesting measurements and searches performed by the ATLAS experiment involve final states with τ-lepton pairs coming from the decay of resonances, including measurements of the properties of the Higgs boson decaying into a pair of τ-leptons (e.g. measurement of the H->ττ coupling, differential cross section measurements and CP measurements), searches for new particles decaying into a pair of τ-leptons (e.g. X->ττ), or searches for new particles decaying into a pair of intermediate particles that can then decay into a pair of τ-leptons (e.g. X->HH->bbττ, X->ZZ->4l, etc.). All these measurements and searches require the reconstruction of the invariant mass of the di-τ system in order to reconstruct the mass of the mother particle and use this information for discriminating signal and background events. Accurate reconstruction of the mass of a resonance decaying into a pair of τ-leptons is very challenging because of the presence of multiple neutrinos from the τ-lepton decays that result in missing transverse energy. Currently, the ATLAS experiment is using a likelihood-based technique to estimate the di-τ mass starting from information on the visible decay products of the τ-leptons and the missing transverse energy. This method provides a mass resolution of about 15% for di-τ masses around 100 GeV, that degrades for higher masses or for Lorentz-boosted systems.

The proposed thesis project will focus on the development of new machine-learning based algorithms to improve the reconstruction of the di-τ mass in terms of improved resolution and improved stability versus the mass values and the Lorentz-boost of the resonance, as well as improved usage of the computational resources.  

Contacts: Alessandra Betti

Study of the first data collected by ATLAS in 2025

Analyses involving muons in the final state are crucial to the ATLAS physics program. Ensuring excellent performance in muon reconstruction, identification, and measurement is essential not only for precision studies of fundamental particles but also for the potential discovery of new physics.

This thesis will focus on the first evaluation of muon reconstruction performance using the new data collected by ATLAS in 2025. By providing real-time feedback on the quality of the data-taking process, this work will play a key role in identifying potential issues and assessing whether the detector and reconstruction algorithms behave as expected. Such early-stage analysis is essential for optimizing performance and ensuring that the collected data meets the requirements for high-precision physics measurements.

Contacts: Luca Martinelli

Entangled in Tops: Quantum Information tests at LHC

Short Abstract: Building on the first experimental observation of top-quark quantum entanglement, the research investigates top-quark-pair correlations with a quantum information approach.

Foreseen Activities: Using ATLAS collision data, the student will develop advanced analysis techniques combining machine learning (ML) for event reconstruction with unfolding methods and statistical inference to extract quantum observables. She/He will search for new physics signatures beyond the Standard Model, pushing the boundaries of quantum measurements at unprecedented energy scales.

Required Skills:  Good knowledge of Python and C++, 
Python libraries for ML and Quantum Information (not strictly required)
Quantum Information Theory (not strictly required)

  First observation of Quantum Entanglement at LHC (Nature 633 (2024) 542)

Contacts: Nello Bruscino, Luca Martinelli 

Precise measurements of the Top Yukawa coupling using advanced machine learning techniques

Short Abstract: The top Yukawa coupling  to the Higgs boson is a crucial parameter to test the Standard Model and search for new physics signs. This thesis will investigate  through Higgs boson production with top-quarks (ttH and tH), with particular focus on the yet-unobserved tH process that controls the relative sign of .

Foreseen Activities: The student will develop a deep-learning algorithm designed to improve identification of prompt leptons from weak boson (Prompt Lepton Isolation Tagger – PLIT). She/He will train and calibrate PLIT, then integrate it into the analysis workflow to enhance signal sensitivity and reduce systematic uncertainties in the top-Higgs coupling measurement.

Required Skills:  Good knowledge of Python and C++, 
Python libraries for Machine Learning (not strictly required)

Contacts: Nello Bruscino, Luca Martinelli, Simonetta Gentile