Proposte Tesi di Laurea Magistrale / Topics for Master Theses 

(the list is indicative: for details contact the teacher directly)

• Topics in High Energy Physics:

• Development and application of new Deep-learning models for applications in High Energy Physics:

  • intrinsically explaianable modular AI based on transformer and fundation models for analysis, interpretation and recasting of HEP analyses at LHC 

  • Large-Language Model based representation learning via multimodal Transformers of physical objects in an High Energy Phsycis Detector (@LHC, @FCCee, ...) and applications to multiboson measurements and searches

    • these two  projects focused in innovative researches, aiming at either developing a new generation of deep learning models that are intrinsecally explainable, based on the recent field of the modular AI, and realizing a digital twin of a collider experiment thta 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. 

    • Tipology: experimental particle physics, artificial intelligence and deep learning 

• Application of state of the art DL models  for the reconstruction and selection of events in high energy physics experiments: 

  • Machine Learning is one of the fields of data analysis and computational research  that has received greater interest and attention in recent years. "Multivariate Analysis", "(Deep) Neural Networks", "Advanced classification or pattern recognition techniques", "Clustering Algorithms" etc., are techniques belonging to a general class of machine learning algorithms that are able to solve in a very efficient and powerful way a great variety of problems in high energy physics ranging from the development of real-time data filter systems (trigger systems), to the offline reconstruction of events, to data analysis proper. Several thesis topics are proposed on the subject, both oriented to the development of new Deep-learning algorithms to be used in the trigger systems of the ATLAS experiment, and to the application of Deep-NeuralNetworks in classification, prediction, anomaly detection, and generative algorithms to be applied in search for new physics and / or precision measurements in the new high-energy and statistical data samples recorded by the ATLAS experiment during Run 2 of LHC and that are currently  acquired in Run 3..

  • Tipology: experimental, computational, modelling, data analysis 

• Development and study of the performances of novel algorithms on FPGAs for muon reconstruction and identification in the ATLAS L0 Muon Barrel Trigger

  • The Level-0 muon trigger system of the ATLAS experiment will undergo a full upgrade for the High Luminosity LHC to stand the challenging requirements imposed by the increase in instantaneous luminosity. The upgraded trigger system will send raw hit data to off-detector processors, where trigger algorithms run on a new generation of FPGAs. To exploit the flexibility provided by the FPGA systems, ATLAS is developing novel algorithms based on both conventional methods and new deep neural network architectures that use quantised weights and activations, optimised to run on FPGAs for efficient reconstruction and identification of muons in the ATLAS "Level-0" trigger. The FPGA represents an optimal solution in this context, because of its flexibility, wide availability of logical resources and high processing speed. In this context the student will work within the ATLAS L0 Muon Barrel trigger group, in the design and deployment of novel fast trigger algorithms based on state-of-the-art machine learning algorithms. The thesis work will include algorithm development and implementation and evaluation of physics performance in terms of efficiency and fake rates and FPGA logic resource occupancy and timing.

  • Tipologia: sperimentale / detector / computazionale-algoritmica

• 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

  • sviluppo di modelli di DNN/RNN/CNN/GNN di ultima generazione per l'identificazione di eventi dovuti a processi di fisica interessanti nel trigger di alto livello (software) dell'esperimento ATLAS. Implementazione degli algoritmi su coprocessori hardware (FPGA/ACAP/GPU) utilizzando librerie software per l'ottimizzazione e la sintesi di nuova generazione (Xilinx VITIS AI, Intel Open VINO) e misura delle prestazioni rispetto ad algoritmi convenzionali.  

  • Tipologia: sperimentale / computazionale-algoritmica  / analisi dati

• Study of novel Deep Graph Network and Transformers models for particle identification and particle flow in the IDEA dual readout calorimeter for future e+e- colliders 

  • Dual-readout calorimetry has emerged as a technique for measuring the properties of high-energy hadrons and hadron jets that offers considerable advantages compared with the instruments that are currently used for this purpose in experiments at the high-energy frontier. The IDEA detector concept for future lepton colliders plan to use a dual-readout fibre calorimeter equipped with fast processors (FPGAs) that allows to implement advanced AI algorithms for reconstruction and particle flow analysis for offline and real-time environments (trigger). The student will work on the design and development of state of the art and novel Deep Learning algorithms based in geometrical deepl learning and in the performance assessment in the context of the identification of different decays of tau leptons and for the particle flow reconstruction. The research works may also includes study of the physics perfromance messured on different benchmark physics processes of interest for the FCCee accelerator program.

  • Tipologia: computazionale-algoritmica-sperimentale / analisi dati 

• Application of DL models at the analysis of test beam data for the prototype of the IDEA dual readout calorimeter  for future e+e- colliders 

  • Partecipation to the data-taking and analysis of the data collected at different test beam facilities for the devlopment of the IDEA dual readout calorimeter. Use of DL models to imporve (real-time and offline) energy and position reconstruction and particle identification

  • Tipologia: sperimenale-detector-computazionale-algoritmica / analisi dati 

• Study of single and di-Higgs production with the ATLAS experiment at LHC

  • According to the Standard Model, the Higgs boson can interact with itself, resulting in the simultaneous production of two Higgs bosons ("di-Higgs production"). This is expected to be a very rare process – so rare that, according to the Standard Model, physicists will need around 20 times more data than available today to be able to measure it. This will be possible with the High-Luminosity upgrade of the LHC. However, physics beyond the Standard Model may enhance the di-Higgs production rate, by altering the strength of the Higgs self-interaction or through an intermediate heavy particle.  The thesis concerns the development and optimisation of a dedicated analysis for the channel HH->bbtautau based on the dataset that ATLAS is collecting in the just started Run-3 of the LHC machine.  In this context different task are possible, ranging from data analysis and optmisation, to trigger studies, to development of machine learning tools. 

  • Tipologia: sperimentale / analisi dati / computazionale

• Search for new physics in di-boson boosted jet topolgies with anomaly detection techniques in the ATLAS experiment

  • In recent years, there has been an explosion of applications of machine learning-based anomaly detection for New Physics searches, enabling the identification of out-of-distribution events that may signal new phenomena or deviations from the Standard Model in a signal-agnostic way. In this research we'll analyze Run-3 data collected by the ATLAS experiment in search for anomalous signature predicted by new physics in final states with multiple boosted jets.  The thesis work includes both experimental analysis of collider detector data and the development of cuting-edge Deep Learning algorithms (GGN, Transformers) for Anomaly detection of fat hadronic jets.

  • Tipologia: sperimentale / analisi dati / computazionale

• Topics in Quantum Computation and Quantum Machine Learning:

  • Sviluppo e implementazione di nuovi algoritmi di Quantum Machine Learning su near-term quantum device (Quantum Anomaly Detection, Quantum-VAE, Quantum-GAN, Quantum Normalizing Flow and Quantum Diffusion Models) e applicazioni in diversi contesti della fisica di base e applicata. 

    • Argomenti: Quantum Computing, Quantum Machine Learning, Differential Programming, Simulazione, Implementazione e ottmizzazione su hardware quantistico

    • Tipologia: computazionale-algoritmica-sperimentale-teorica, also togheter industrial partners 

  • Sviluppo e implementazione di algoritmi di Deep Learning Classico per la preparazione, monitoring e ottmizzzazione di computer quantistici: Transpiling, Quantum Noise simulation, Qubit error correction, ...

    • Argomenti: Quantum Computing, Machine and Deep Learning

    • Tipologia: computazionale-algoritmica-sperimentale-teorica, also togheter industrial partners 

  • Sviluppo di algoritmi quantistici per Quantum sensing e QML su quibit analogici superconduttivi 

  • Sviluppo di algoritmi di QML su circuiti fotonici

    • Argomenti: Quantum Computing, Machine and Deep Learning

    • Tipologia: computazionale-algoritmica-sperimentale-teorica

1. Topics in foundaional of applied Machine Learning, Deep Learning, AI and Physics:

• Artificial Intelligence for non-linear spectroscopy (insieme al gruppo di spettroscopia Raman del Prof. T. Scopigno)

  • La spettroscopia Raman non-lineare è uno strumento potente in grado di indagare le proprietà vibrazionali della materia su scala atomica. Rispetto alla spettroscopia Raman convenzionale, presenta numerosi vantaggi, come la possibilità di mappare fenomeni ultraveloci e determinare il moto atomico su scale temporali del femtosecondo.  Tuttavia, a causa della non linearità dei processi ottici su cui si basa, i segnali misurati sperimentalmente sono generalmente più complessi e difficili da interpretare: le forme di righe possono apparire sovrapposte ad un segnale di background e subire forti distorsioni, con profili spettrali che possono apparire sia negativi che dispersivi. La tesi ha carattere teorico-sperimentale, prevede infatti di sviluppare tramite algoritmi di Intelligenza Artificiale una Rete Neurale in grado di isolare il segnale Raman dai contributi non-lineari spuri. Gli algoritmi verranno allenati mediante opportuni training set ottenuti dal calcolo delle risposte Raman non-lineari per sistemi modello. Come benchmark per la validazione del protocollo sviluppato saranno utilizzati dati acquisiti sperimentalmente su sistemi reali

  • Tipologia: sperimentale / laboratorio / computazionale-algoritmica 

• Studio di algoritmi di Deep Learning basati su Graph Neural Network per la classificazione e la ricomposizione di oggetti solidi divisi in frammenti (in collaborazione con l'opificio delle pietre dure di Firenze)

  • La ricostruzione di oggetti solidi divisi in frammenti o regioni di dimensioni e caratteristiche diverse è un problema non risolto da lungo tempo, che ha interessato le comunità di grafica computazionale, matematica, e più recentemente del machine learning. A differenza del caso 2D la complessità maggiore del caso 3D fa si che non esistano ad oggi tecniche generali consolidate per affrontare queste casistiche. Negli ultimi anni in particolare l’applicazione di algoritmi di deep learning a questi problemi su oggetti 3D ha portato a risultati significativi in campi molto diversi, che vanno dalla medicina, alla fisica fino all’ambito dei beni culturali. Si veda a tal proposito l’applicazione degli algoritmi di deep learning alla catena di ricostruzione dell’Esercito di Terracotta (inserito nell’elenco dei patrimoni dell’umanità dell’Unesco). Si propone una tesi di carattere teorico-sperimentale che prevede lo sviluppo di algoritmi innovativi di deep learning basati su Graph Neural Networks per la ricostruzione di un oggetto a partire dai suoi frammenti. In una prima fase verrà sviluppato il modello teorico, ed in seguito l’algoritmo verrà applicato ad uno o più casi reali nel contesto della ricostruzione di manufatti antichi.

  • Tipologia: computazionale-algoritmica 

• Applicazioni avanzate di Machine Learning e Deep Learning in fisica medica

  • Argomenti: sviluppo di nuovi algoritmi e applicazioni in imaging medico (MRI, CT-Scan, X-ray), applicazioni in radioterapia (Tomotherapy QA), applicazioni in diagnostica/prognostica medica, sviluppo di nuovi algoritmi per la correzione di movimento e il denoising di immagini MRI in fluoro 19. Studio e applicazione di modelli generativi di AI per il drug discovery.

  • Tipologia: computazionale-algoritmica / analisi dati / sviluppo dimostratori

• Development of novel physics inspired Deep Learning and AI explainability models

  • Arguments: development of physics inspired Deep Learning model to simplify and enhance transparency and interpretability of machine learning models

  • Tipologia: computazionale-algoritmica / machine learning / fundational

• Sviluppi teorici in topological quantum field theory e sviluppo di nuovi algoritmi per l'encoding topologico di informazioni basati su geometrical deep learning (graph neural network equivarianti per trasformazioni fi gauge)

  • Argomenti: QTFT, spin networks e materiali topologici, Quantum Neural Network, Deep Learning, Equivariant Graph Neural Networks

  • Tipologia: teorica (teoria dei campi) o computazionale-algoritmica (QNN/DNN)

• Sviluppo di nuovi odelli basati su Graph Neural Network ispirati alla fisica per la soluzione di problemi di ottmizzazione combinatoriale (Graph Coloring, problemi sat, etc.)

  • The graph coloring problem is an optimization problem involving the assignment of one of q colors to each vertex of a graph such that no two adjacent vertices share the same color. This problem is NP-hard and arises in various practical applications. We'll develop and test  novel algorithms that leverages graph neural networks to tackle the problem efficiently, particularly for large graphs. The approach is physics-inspired, an approach that leverages tools used in statistical mechanics to improve the training and performance of the algorithm

  • Tipologia: teorico-computazionale-algoritmica

• Sviluppo di nuovi modelli basati su Graph Neural Network basati su PU-learning per l'analisi di knowledge graph di grande dimensione

  • nderstanding and predicting relationships within complex systems is a key challenge across many disciplines. Systems like ecosystems, social networks, and climate models resist full characterization due to their complexity and numerous variables. When system dynamics and interactions between components are only partially understood, machine learning becomes a vital tool for uncovering patterns and hidden relationships from data. However, real-world datasets often face limited availability and high acquisition costs, capturing only a fraction of possible interactions. These datasets are also prone to sampling biases influenced by practical or economic constraints, leading to incomplete and non-i.i.d. (independent and identically distributed) data. This challenges traditional machine learning methods, which typically require complete and unbiased datasets. This thesis proposes developing algorithms that work without fully supervised data. The goal is to create a theoretical framework for machine learning methods that analyze complex systems with interactive components, even when data is biased and incomplete.

  • Tipologia: teorico-computazionale-algoritmica

Proposte tesi Laurea Triennale: contattare il docente