Posts by Collection

portfolio

publications

The Classical Solution for the Bosonic String in the Presence of three D-branes Rotated by Arbitrary SO(4) Elements

Published in Nucl. Phys. B, 941, 158-194, 2019

We consider the classical instantonic contribution to the open string configuration associated with three D-branes with relative rotation matrices in SO(4). Read more

Cite as: R. Finotello, I. Pesando. 'The Classical Solution for the Bosonic String in the Presence of three D-branes Rotated by Arbitrary SO(4) Elements'. Nucl. Phys. B 941 (2019), pp. 158-194. https://doi.org/10.1016/j.nuclphysb.2019.02.010

HyperPCA: a Powerful Tool to Extract Elemental Maps from Noisy Data Obtained in LIBS Mapping of Materials

Published in Spectrochim. Acta B: At. Spectrosc., 192, 106418, 2022

We introduce HyperPCA, a new analysis tool for hyperspectral images to consistently reconstruct the spectroscopic signal, with a particular emphasis on LIBS data. Read more

Cite as: R. Finotello, M. Tamaazousti, J.-B. Sirven. 'HyperPCA: a Powerful Tool to Extract Elemental Maps from Noisy Data Obtained in LIBS Mapping of Materials'. Spectrochim. Acta B: At. Spectrosc., 192 (2022), 106418. https://doi.org/10.1016/j.sab.2022.106418

Trustworthiness of Laser-Induced Breakdown Spectroscopy Predictions via Simulation-based Synthetic Data Augmentation and Multitask Learning

Published in arXiv, 2022

We build an end-to-end pipeline to deal with data augmentation, robust multitask deep calibration models and the trustworthiness assessment of libs predictions Read more

Cite as: R. Finotello, D. L'Hermite, C. Quéré, B. Rouge, M. Tamaazousti, J.-B. Sirven. 'Trustworthiness of Laser-Induced Breakdown Spectroscopy Predictions via Simulation-based Synthetic Data Augmentation and Multitask Learning'. arXiv:2210.03762. https://arxiv.org/abs/2210.03762

Deep Learning Complete Intersection Calabi–Yau Manifolds

Published in Machine Learning in Pure Mathematics and Theoretical Physics, pp. 151-181 (2023), 2023

We review advancements in deep learning techniques for complete intersection Calabi–Yau (CICY) 3- and 4-folds. Read more

Cite as: H. Erbin, R. Finotello, 'Deep Learning Complete Intersection Calabi–Yau Manifolds', in Machine Learning in Pure Mathematics and Theoretical Physics, pp. 151-181 (2023), World Scientific, edited by Y.-H. He. https://doi.org/10.1088/2632-2153/ac37f7

Interlaboratory comparison for quantitative chlorine analysis in cement pastes with laser induced breakdown spectroscopy

Published in Spectrochim. Acta B, 202, 106632, 2023

This paper presents the results of an interlaboratory comparison on the analysis of Cl in cement paste samples conducted by 12 laboratories in 10 countries. Read more

Cite as: T. Völker et al. 'Interlaboratory comparison for quantitative chlorine analysis in cement pastes with laser induced breakdown spectroscopy'. Spectrochim. Acta B, 202, 106632. https://doi.org/10.1016/j.sab.2023.106632

Deep Learning Complete Intersection Calabi–Yau Manifolds

Published in Machine Learning in Pure Mathematics and Theoretical Physics, 2023

We review advancements in deep learning techniques for complete intersection Calabi–Yau (CICY) 3- and 4-folds. Read more

Cite as: H. Erbin, R. Finotello, Deep Learning Complete Intersection Calabi–Yau Manifolds. In Machine Learning in Pure Mathematics and Theoretical Physics, World Scientific, pp. 151-181 (2023), edited by Y.-H. He https://doi.org/10.1142/9781800613706_0005

talks

Exploring Particle Physics in 2D BCFT: D-branes, Twist Fields and Defect CFT

Published:

In this short seminar for the midterm of the Ph.D. programme, we explore topics on particle physics seen from the perspective of 2D BCFT and open strings. We build a Standard Model-like scenario with semi-phenomenological properties using intersecting D-branes in $\mathrm{SO}(4)$ at angles and time-dependent defects. We then study the properties of the arising non Abelian twist fields and Abelian spin fields. Read more

Spin Fields as Point–like Defects on the Worldsheet

Published:

We show a new method to compute the correlator of an arbitrary number of (excited) spin fields based ona time dependent defect CFT procedure, with the possibility to extend it to (excited) twist fields, both in the Abelian and non Abelian cases Read more

Time Dependent Defect CFT and Excited Spin Fields

Published:

In this video-poster, presented online due to the COVID-19 pandemic, we present a new method to compute amplitudes in the presence of excited spin fields, without resorting to bosonisation. We introduce point-like defects on the open superstring worldsheet, and we show that, though time dependent, the theory is still conformal. Read more

Inception Neural Networks for Calabi-Yau Manifolds

Published:

In this talk, still online due to the COVID-19 pandemic, we introduce a new neural network architecture, inspired by Google’s Inception network, capable of predicting the Hodge number $h^{1,1}$ with more than $99\%$ of accuracy. The results represent a large improvement on previous attempts at the same task, by increasing the reached accuracy by more than $20\%$. Read more

An AI Perspective on Phenomenology and Strings

Published:

Particle physics and cosmology are fascinating topics both from an experimental point of view and a theoretical perspective. The balance between theoretical predictions and physical evidence is however delicate and usually requires contributions from several research areas. In the effort to find a unified description of fundamental forces of nature, the framework of String Theory provides some of the needed tools. We present how hints of experimental evidence of particle physics might be recovered using different mathematical models, and how cosmological singularities such as the Big Bang can find some explanations inside a string theory. Read more

Algebraic Geometry and Computer Vision: Inception Neural Network for Calabi-Yau Manifolds

Published:

Computing topological properties of Calabi-Yau manifolds is, in general,a challenging mathematical task: traditional methods lead to complicated algorithms, without expressions in closed form in most cases. At the same time, recent years have witnessed the rising use of deep learning as a method for exploration of large sets of data, to learn their patterns and properties. This is specifically interesting when it comes to unravel complicated geometrical structures, as it is a central issue both in mathematics and theoretical physics, as well as in the development of trustworthy AI methods. Motivated by their distinguished role in string theory for the study of compactifications, we try to compute the Hodge numbers of Complete Intersection Calabi-Yau (CICY) 3-folds using deep neural networks. We focus on architectures involving convolutional layers, as most modern applications (both in research and in the industry) profit from having the shared parameters for feature creation and recognition of patterns in the input. As such, we map the original task to a computer vision problem, reminiscent of object identification. We introduce a new regression neural network, inspired by Google’s Inception network, which leverages the theoretical knowledge on the inputs, with the recent advancements in AI. As a result, we reach 97\% of accuracy in the prediction of h1,1 with just 30\% of the available data for training, and almost perfect accuracy with 80\% training ratio, outperforming by a large margin previous results. This shows the potential of deep learning to learn from geometrical data, and it proves the versatility of architectures developed in different contexts, which may therefore find their way in theoretical physics and mathematics for exploration and inference. Read more

Computer Vision and Algebraic Geometry: AI for Theoretical Physics

Published:

We rephrase a central problem in algebraic geometry and theoretical physics as a computer vision task. Leveraging knowledge of the physical data with the introduction of architectures inspired by Google’s Inception network, we reach almost perfect accuracy on the predictions. We thus prove the versatility of the AI models and their reliability in making accurate physical predictions. Read more

Algebraic Geometry and Computer Vision: Inception Neural Network for Calabi-Yau Manifolds

Published:

Computing topological properties of Calabi-Yau manifolds is, in general, a challenging mathematical task: traditional methods lead to complicated algorithms, without expressions in closed form in most cases. At the same time, recent years have witnessed the rising use of deep learning as a method for exploration of large sets of data, to learn their patterns and properties. This is specifically interesting when it comes to unravel complicated geometrical structures, as it is a central issue both in mathematics and theoretical physics, as well as in the development of trustworthy AI methods. Motivated by their distinguished role in string theory for the study of compactifications, we compute the Hodge numbers of Complete Intersection Calabi-Yau (CICY) manifolds using deep neural networks. Specifically, we introduce new regression architectures, inspired by Google’s Inception network and multi-task learning, which leverage the theoretical knowledge on the inputs with recent advancements in AI. This shows the potential of deep learning to learn from geometrical data, and it proves the versatility of architectures developed in different contexts, which may therefore find their way in theoretical physics and mathematics for exploration and inference. Read more

Sparse Representations and Kernel-based PCA: Powerful Tools to Extract Elemental Maps from Noisy Data Obtained in LIBS Mapping of Materials

Published:

Laser-induced breakdown spectroscopy (LIBS) is a preferred technique for fast and direct multielemental mapping of solid samples under ambient pressure, with a lateral resolution of a few µm. LIBS mapping data have two peculiarities: an intrinsically low signal-to-noise ratio due to single-shot measurements, and a high dimensionality due to the elevated number of spectra acquired. The extraction of physico-chemical information from a noisy and large dataset is therefore a major issue. Multivariate approaches were introduced as a means to cope with such data, particularly using Principal Component Analysis (PCA). Yet, PCA is known to present theoretical constraints to the consistent reconstruction of the dataset, and it is therefore limited to efficiently interpret LIBS mapping data. We introduce a new analysis strategy, called DWT-kPCA, based on a sparse representation of the data using Discrete Wavelet Transform and kernel-based sparse PCA to remove noise from the data and consistently reconstruct the spectroscopic signal. We show that the method presents advantages both in quantity and in quality of the information recovered, thus improving the physico-chemical characterisation of analysed surfaces. Read more

Applying Machine Learning to String Theory

Published:

In these lectures, we present an introduction to neural networks and their interest for physics and mathematics. We also show several relevant applications of deep learning in the field of theoretical physics. More information can be found in the GitHub repository. Read more

Deep Multi-task Mining Calabi-Yau Manifolds

Published:

Computing topological properties of Calabi-Yau manifolds is a challenging mathematical task. Recent years have witnessed the rising use of deep learning as a method for exploration of large sets of data, to learn their patterns and properties. This is specifically interesting when it comes to unravel complicated geometrical structures, as well as in the development of trustworthy AI methods. Motivated by their distinguished role in string theory for the study of compactifications, we compute the Hodge numbers of Complete Intersection Calabi-Yau manifolds using deep neural networks. Specifically, we introduce a regression architecture, based on GoogleNet and multi-task learning, capable of mining information to produce highly accurate simultaneous predictions. This shows the potential of deep learning to learn from geometrical data, and it proves the versatility of architectures developed in different contexts. Read more

HyperPCA. Une méthode d’analyse innovante pour l’imagerie hyperspectrale

Published:

Contexte

Le projet CAMELIA (CArtographie Multi-Élémentaire par LIBS et Intelligence Artificielle, PTC-ID 2021-2022) porte sur la cartographie multi-élémentaire à l’échelle micrométrique d’échantillons par spectroscopie de plasma induit par laser (« Laser-Induced Breakdown Spectroscopy », ou LIBS, en anglais). Cette technique utilise un faisceau laser focalisé sur la surface d’un échantillon pour créer un plasma dont l’émission est caractéristique des éléments présents. Des cartographies en deux dimensions de la surface sont obtenues en déplaçant l’échantillon à chaque tir du laser : chaque pixel de la cartographie correspond à un spectre, qui contient l’information sur les éléments présents dans le plasma d’ablation. Cependant, cette technique a un rapport signal/bruit intrinsèquement faible, dû à l’utilisation d’un seul tir laser par cratère pour augmenter la résolution latérale. Elle présente également une dimensionnalité très élevée, liée au nombre de cratères nécessaires pour cartographier une surface donnée. Par conséquent, l’extraction de l’information physico-chimique de ces données fortement bruitées et de grande dimension, est un enjeu majeur.

Objectifs

Dans CAMELIA, on développe des techniques basées sur l’apprentissage automatique et l’intelligence artificielle, pour exploiter efficacement le signal. L’objectif est de mettre au point une méthode pour s’affranchir au maximum du bruit des données et pour extraire le signal spectroscopique. Cette technique doit permettre la reconstruction de la distribution des éléments chimiques sur la surface de l’échantillon. Par ailleurs, la méthodologie doit être interprétable sur le plan spectroscopique et physico-chimique, et non supervisée, car aucune connaissance a priori des échantillons ne doit être requise.

Méthodologie

Dans le domaine de la cartographie LIBS, outre la méthode usuelle d’exploitation des spectres à partir l’intensité de raies prédéfinies, l’Analyse en Composantes Principales (ACP) a été proposée par plusieurs auteurs car elle permet de fournir des résultats interprétables et de façon non supervisée [1]. Cette méthode permet de cartographier en deux dimensions la distribution des éléments chimiques. La méthodologie proposée est basée sur des développements récents de l’ACP [2], qui ont montré des résultats prometteurs pour la réduction du bruit et la reconstruction du signal. Pour résoudre les problèmes soulevés par la spécificité des données de cartographie LIBS, nous avons proposé dans CAMELIA la technique HyperPCA [3, 4], basée sur le couplage d’une transformée en ondelettes pour la création d’une représentation parcimonieuse des données et la résolution des interférences spectrales, et d’une approche basée sur une fonction noyau pour la réduction du bruit. Cette procédure permet de s’affranchir de la présence d’une distribution du bruit aléatoire pour des jeux de données de grande dimensionnalité : cette configuration est typique de la cartographie LIBS et de l’imagerie hyperspectrale, ce qui permet l’exploitation de cette méthode. Notre démarche a donc consisté à comparer les résultats obtenus par l’approche univariée usuelle, par l’ACP et par l’HyperPCA, sur des jeux de données simulées et expérimentales.

Résultats

L’HyperPCA fournit une plus grande quantité d’information, avec une meilleure qualité, par rapport à la méthode usuelle et par rapport à l’ACP : on obtient un grand nombre de composantes lisibles même en présence d’un rapport signal/bruit très faible. De plus, l’utilisation d’une transformée en ondelettes permet de capturer les propriétés physiques des profils des raies d’émission. On observe enfin que les composantes calculées par HyperPCA sont souvent mono-élémentaires, ce qui permet d’obtenir des cartographies plus facilement interprétables.

Conclusion et perspectives

L’HyperPCA a été introduite dans le projet CAMELIA pour l’analyse des données hyperspectrales (LIBS en particulier). L’algorithme proposé montre des améliorations significatives par rapport aux méthodes de l’état de l’art, pour l’extraction de l’information physico-chimique, puisqu’il peut être employé en présence d’un rapport signal/bruit très faible. La quantité d’information récupérée et la qualité des composantes principales sont les avantages les plus évidents de cette technique. L’HyperPCA permet de grouper les raies d’émission dans les composantes par élément chimique, ce qui donne des cartographies facilement interprétables en termes de contributions mono-élémentaires. En perspective, on peut envisager d’appliquer cette approche à l’analyse quantitative par LIBS, car elle offre une méthode de réduction de dimensionnalité et d’extraction de composantes de très bonne qualité. Il serait également intéressant d’explorer d’autres applications de l’HyperPCA pour des tâches de segmentation d’image, ou encore d’étudier des méthodes plus avancées, comme l’ACP tensorielle, pour tenir compte de la distribution spatiale des données dans les cartographies.

Références

  • [a] R. Finotello, M. Tamaazousti and J.-B. Sirven, « Méthodes d’analyse en composantes principales innovantes pour l’imagerie hyperspectrale », Séminaire des 60 ans de la CETAMA, 19 – 21 octobre 2021, Nîmes, France
  • [b] R. Finotello, M. Tamaazousti and J.-B. Sirven, « HyperPCA: An Advanced Framework of Principal Components Analysis for Hyperspectral Images », Séminaire annuel PE-PTC, 22 – 24 novembre 2021, Grenoble, France
  • [c] R. Finotello, M. Tamaazousti and J.-B. Sirven, « Sparse Representations and Kernel-based PCA », EMSLIBS 2021, 29 novembre – 2 décembre 2021, Gijón, Espagne
  • [1] L. Jolivet et al., “Review of the recent advances and applications of LIBS-based imaging,” Spectrochimica Acta Part B: Atomic Spectroscopy 151 (2019) 41–53. doi:10.1016/j.sab.2018.11.008.
  • [2] M. E. A. Seddik, M. Tamaazousti, and R. Couillet, “A kernel random matrix-based approach for sparse PCA,” 2019. https://openreview.net/forum?id=rkgBHoCqYX
Read more

Computer Vision for Physics. Theory and Experiments

Published:

Ces dernières années ont été marquées par un nombre toujours croissant d’applications de techniques basées sur l’apprentissage automatique à différents domaines théoriques et expérimentaux. Les progrès des algorithmes non supervisés ont permis l’exploration de certaines structures invisibles, tandis que les progrès de l’apprentissage profond ont rendu possible l’analyse de systèmes complexes. Un exemple de cette incroyable évolution est l’aptitude des ordinateurs à effectuer des tâches de détection et de segmentation d’objets imitant la vision humaine. Récemment, l’étude du comportement des algorithmes, que l’on trouve typiquement dans le cadre de la vision par ordinateur, a ouvert la possibilité d’applications dans les analyses qualitatives et quantitatives en physique. Dans ce séminaire, nous montrons le lien entre la théorie sous-jacente à certaines techniques supervisées et non supervisées et leur application aux données de la physique expérimentale. Nous préparons d’abord le terrain avec une application de vision par ordinateur de l’apprentissage profond à un problème mathématique afin de mieux illustrer les possibilités de l’intelligence artificielle. Nous passons ensuite à l’analyse de données expérimentales, avec une attention particulière pour la spectroscopie du plasma induit par laser. En détail, nous traitons de l’extraction non supervisée de cartographies élémentaires en présence d’un fort bruit de fond, et de la quantification de la confiance dans l’apprentissage profond en utilisant des réseaux de neurones multitâches pour l’étalonnage supervisé. Read more

Helping AI Understand Physics. Trustworthy Approaches to Hyperspectral Imaging

Published:

Can trust be measured? Does a computer always know what to do? When dealing with experimental measurements of any kind, the ability to assess stability and trustworthiness of a machine learning model is key to create efficient analysis tools. In the case of hyperspectral data, knowledge of the physics underlying their generation is an inductive bias useful to assess the confidence of quantitative predictions. However, the reconstruction of a hyperspectral signal is often prevented by the presence of noise and strong spectral interference. Recent developments show that, under sensible assumptions, mathematical transformations of the data can ease the automatic extraction of information in complicated situations. In this talk, we review the state-of-the-art in applications of near-infrared and laser-induced breakdown spectroscopy, with a specific focus on trustworthy AI issues and semantic segmentation of hyperspectral images. Read more

Machine Learning for Complete Intersection Calabi-Yau Manifolds

Published:

We describe the recent developments in using machine learning techniques to compute Hodge numbers of complete intersection Calabi-Yau (CICY) 3- and 4-folds. The main motivation is to understand how to study data from algebraic geometry and solve problems relevant for string theory with machine learning. We describe the state-of-the art methods which reach near-perfect accuracy for several Hodge numbers, and discuss extrapolating from low to high Hodge numbers, and conversely. Read more

Trustworthiness of Laser-Induced Breakdown Spectroscopy Predictions via Simulation-based Synthetic Data Augmentation and Multitask Learning

Published:

Laser-induced breakdown spectroscopy is a technique able to perform fast measurements in ambient air without any limitation on the targeted elements. It consists in focusing a laser beam on the surface of a sample in order to form a plasma and in analyzing the radiation emanating from its cooling. It is a versatile procedure used in various scenarios, such as nuclear decommissioning or fundamental physics experiments, for qualitative and quantitative spectral analyses. For the latter, the objective is usually to build a model relating experimental spectra to the concentration of the species of interest. This is based on the availability of a calibration set of known samples and can be done through a variety of supervised techniques. However, in the most straightforward implementation, models do not estimate to which extent an unknown sample is well represented by the calibration set. Hence, we do not know, in general, how reliable the prediction is. For this purpose, we build robust calibration models using deep convolutional multitask learning architectures to predict the concentration of the analyte, alongside additional spectral information as auxiliary outputs. Due to the experimental lack of training samples, we introduce a simulation-based data augmentation process to synthesize an arbitrary number of spectra for training, statistically representative of the experimental data. The secondary predictions are finally used to validate the model’s trustworthiness by taking advantage of the mutual dependencies of the parameters of the multitask neural networks: a statistical analysis of the outputs can be directly performed through a comparison with ground truth quantities. Such an end-to-end pipeline has a good ability in detecting anomalies and out-of-distribution samples without the need for a separate elemental analysis. Results on different types of materials, such as cement samples and alloys, show an improvement in the robustness (seen as homoscedasticity) and the trueness of the predictions, especially in the presence of noise and strong spectral interference in the spectra. Read more

HyperPCA : à l’interface entre la théorie des matrices aléatoires et la spectroscopie du plasma induit par laser

Published:

La spectroscopie du plasma induit par laser (« Laser-Induced Breakdown Spectroscopy », ou LIBS, en anglais) est une technique pour la cartographie multiélémentaire rapide et directe d’échantillons sous pression ambiante, sans aucune limitation quant à l’élément ciblé. Les données représentent des images hyperspectrales de la surface d’un échantillon : chaque pixel est un spectre d’émission élémentaire, où le nombre de canaux peut être très important (1k à 100k canaux). Cependant, les données de cartographie LIBS présentent deux particularités : un rapport signal/bruit intrinsèquement faible en raison des mesures uniques, et une dimensionnalité élevée en raison du grand nombre de spectres acquis pour l’imagerie. Ceci est d’autant plus vrai que la résolution latérale augmente : dans ce cas, le diamètre du spot d’ablation est réduit, de même que la masse ablatée et le signal d’émission, tandis que le nombre de spectres pour une surface donnée augmente. Par conséquent, l’extraction efficace d’informations physico-chimiques à partir d’un ensemble de données bruitées et volumineuses est un problème majeur. Des approches multivariées ont été introduites par plusieurs auteurs comme moyen de traiter de telles données, en particulier l’analyse en composantes principales (« Principal Components Analysis », ou PCA, en anglais) [1-2]. Cette technique est utile pour analyser les corrélations entre différents éléments, mais elle est limitée aux faibles rapports signal/bruit. Read more

teaching

Physics at LHC

Workshop, STEM Workshop for High School students, 2015

The series of lectures on accelerator physics (with emphasis on the LHC at CERN, Geneva) is aimed at Italian high school students as complementary preparation in physics and mathematics. The lectures were given in English. Read more

Waves, Fluids and Thermodynamics

Undergraduate course, Università degli Studi di Torino, 2019

I held the tutoring and exercise sessions for 1st year students during the final trimester of the Bachelor of Science in Physics. Read more

Physics 1

Undergraduate course, Università degli Studi di Torino, 2020

I held the tutoring and exercise sessions for 1st year students during the final semester of the Bachelor of Science in Physics. Read more

Internships

Internships, Commissariat à l'énergie atomique et aux énergies alternatives, 2023

Supervision of students during their internships. Read more