En utilisant ce site, vous acceptez que les cookies soient utilisés à des fins d'analyse et de pertinence     Oui, j'accepte  Non, je souhaite en savoir plus

Coronavirus - A partir du lundi 16/03/2020, l'INSTN est fermé jusqu’à nouvel ordre Informations

Sujets de thèse
Filtrer par critères

DRF : Sujet de thèse SL-DRF-20-0650

Physique théorique / Physique théorique
INTITULÉ DU SUJET Français English

Apprentissage automatique en physique fondamentale


Context: Today, as a result of an outstanding experimental program, particle physics and cosmology are

flooded with data. This offers us the unique opportunity to answer questions that have been the object of

speculation for more than fifty years. In particular, particle colliders and Cosmic Microwave Background

(CMB) surveys have collected a wealth of high-quality data in the past decades and may continue to do so for

decades to come. Analyzing these datasets effectively is key to make progress in fundamental physics.

In particle physics, as more and more data are collected, the problems that confront us become sharper

and harder to solve. We know that the theories that well describe current data are incomplete and should be

extended, but our prior beliefs on how the extension should look like and on where to discover it experimentally

become less concordant every day. This calls for a model-independent approach to data analysis.

In cosmology, well established measurements, such as the rate of expansion of the Universe from the CMB

radiation, are being challenged by new observations. Reconciling these different observations or understanding

if they signal the presence of new physical phenomena, can only be done by interrogating the data in novel

ways. Mining these large, multivariate datasets for signs of new phenomena, presents many of the challenges

that machine learning and deep learning are overcoming in a variety of other domains. In this project we

leverage the unprecedented growth of these fields to shed light on the open questions of fundamental physics.

Thesis Project: We consider the problem of having large multivariate datasets that are seemingly well de-

scribed by a reference model. Departures from the reference model can be statistically significant, but are

caused only by a very small fraction of events. The significance might stem from the extreme rarity of the

discrepant events in the reference model and in this case anomaly detection techniques might be employed.

Or the discrepancy is due to a small excess (or even a deficit) of events in a region of the space of physical

observables that is also populated in the reference model. Our goal is to determine if the experimental dataset

does follow the reference model exactly or if it instead contains “smal” departures as described above. In the

latter case, we also want to know in which region of the space of observables the discrepancy is localized. This

problem is relevant to Large Hadron Collider (LHC) datasets that are well described by the Standard Model of

particle physics (SM) and CMB datasets that are well described by the standard cosmological model ?CDM.

I have already developed a new machine learning technique [1] that allows to analyze large datasets in

a model-independent way, detecting data departures from a given reference model with no prior bias on the

nature of the new physics responsible for the discrepancy. There are a number of potential applications in

particle physics, astrophysics and cosmology. We are already taking the relevant steps in collaboration with

CMS experimentalists to use this technique on LHC data. The next step is to apply this technique to CMB

datasets in order to detect deviations from ?CDM (which can contribute to shed light on the current tension

between different measurements of the Hubble parameter). The application of this technique to CMB data will

be the main focus of the project.

In parallel we will refine [1] and look for the optimal model-independent new physics detection strategy.

This problem can be reduced to finding the minimum of a suitable loss function. Furthermore, the loss function

first developed in [1] can be used to obtain better performances in traditional classification problems, with a

variety of possible applications to fundamental physics (such as for example the separation between quark and

gluon jets). There are many other directions that are worth exploring such as embedding [1] in a Generative

Adversarial Network and using Autoencoders to compute a p-value rather than just signaling the presence of

anomalous events. Depending on the progress of the student during his thesis project we will explore all these

directions or only a subset, with the priorities outlined above: 1) CMB data mining 2) Optimal search strategy

3) Applications in classification 4) Variations on Autoencoders.

[1] R. T. D’Agnolo and A. Wulzer, Learning New Physics from a Machine, Phys. Rev. D 99, no. 1, 015014

(2019) doi:10.1103/PhysRevD.99.015014 [arXiv:1806.02350 [hep-ph]].


M2 physique theorique

Institut de Physique Théorique
Service de Physique Théorique
Centre : Saclay
Date souhaitée pour le début de la thèse : 01/10/2020

Raffaele D´Agnolo  

Liste des pôles/Liste des départements/Liste des services/SPhT
91191 Gif sur Yvette

Téléphone : +33 1 69 08 66 30

Physique en Île-de-France (EDPIF)