Physics in the AI era

24 Sept 2024, 08:00 → 27 Sept 2024, 19:30 Europe/Rome

University of Pisa

WELCOME TO PHYSICS IN THE AI ERA - An initiative by Università di Pisa

Young Organization for Young Researchers 

Physics in the AI era ($\Pi$AIe) is the first International Conference completely organized by Young Researchers which supports the active participation of Young Researchers.

The event starts from an initiative of Università di Pisa and it is organized to provide a balanced alternation of "Senior" and "Young" talk sessions on the field of the Artificial Intelligence and Physics. This will give the opportunity to early-stage researchers, severely penalized by the pandemic, to spread their work, but also to experienced scientists to be aware of original research inputs coming from younger physicists.

The workshop will involve the Department of Physics at the University of Pisa and the Scuola Normale Superiore. Furthermore, it will engage both the theoretical and experimental research groups, both of which are large, active, and rich in personalities in Pisa. This will allow many more PhD students and young researchers to participate and gain productive experiences from the workshop.

In recent years, artificial intelligence (AI) has revolutionized numerous scientific and technological sectors. However, in the field of physics, the potential for interdisciplinary dialogue is often overlooked.

This workshop aims to bridge this gap by promoting constructive dialogue between the physics community and that of artificial intelligence. The idea is to explore together the potential that AI can offer in the field of physics and how this can bring significant benefits to both communities, inviting renowned experts from both sectors.

The workshop will take place over 4 days, three of which will focus on a theme currently central in scientific discourse, in particular regarding Complex Systems Physics, Particle Physics, and Cosmology. The days will be organized so that in the morning, experts in computer science will describe cutting-edge techniques related to AI, while in the afternoon, expert physicists and young researchers will introduce how these techniques can be utilized in their field.

IMPORTANT NOTE: for every issue  regarding the conference, do not rely on emails accounts other than the official (piaie@lists.pi.infn.it) and the organizers' ones!

No FEE is required.

THE WORKSHOP IS SUPPORTED BY

Fellowship_postdoc.docx

Fellowship_student.docx

Tuesday, 24 September

16:00 Registration, welcome drink and poster session

Wednesday, 25 September

Cosmology

1 AI in 21cm cosmology overview

Speaker: Prof. Andrei Mesinger (Scuola Normale Superiore)

2 Plenary talk

Speaker: Dr Caroline Heneka (University of Heidelberg)

3 Plenary talk

Speaker: Dr David Prelogovic

10:30 Coffee break

10:30 Poster Session

Cosmology

4 Plenary Talk

Speaker: Prof. Kana Moriwaki (Univeristy of Tokyo)

5 Machine learning based inference of high redshift observations.

High redshift observations, such as the 21cm signal, luminosity functions, etc., carries immense potential to probe the formation and properties of the first galaxies, and beyond $\Lambda \mathrm{CDM}$ Cosmology. However, a complete statistical analysis of such observation is limited by the time consuming nature of simulators. Using machine learning based emulators we can overcome this obstacle, and produce fast and accurate realizations of these observables. Here we present two applications of this approach: (i) reproducing HERA constraints on X-ray luminosity in the early universe from the 21cm power spectrum, and re-evaluating these bounds in the presence of PopIII stars. (ii) Achieving new constraints on Fuzzy Dark Matter using luminosity functions, the Thomson scattering optical depth of CMB photons, and upper bounds on the neutral fraction at $z\sim 6$. In addition, we forecast that upcoming observations of the 21cm power spectrum can improve these bounds.

Speaker: Hovav Lazare (Ben Gurion University)

6 Evaluating Summary Statistics with Mutual Information for Cosmological Inference

The ability to compress observational data and accurately estimate physical parameters relies heavily on informative summary statistics. In this paper, we introduce the use of mutual information (MI) as a means of evaluating the quality of summary statistics in inference tasks. MI can assess the sufficiency of summaries, and provide a quantitative basis for comparison. We propose to estimate MI using the Barber-Agakov lower bound and normalizing flow based variational distributions. To demonstrate the effectiveness of our approach, we conduct a comparative analysis of three summary statistics: the power spectrum, bispectrum, and scattering transform. Our comparison is performed within the context of inferring physical parameters from simulated CMB maps and highly non-Gaussian 21cm mock observations. Our results highlight the ability of our approach to correctly assess the informativeness of different summary statistics, enabling the selection of an optimal set of statistics for inference tasks.

Speaker: Ce Sui

12:15 Discussion - AI in 21-cm Cosmology

13:00 Lunch break

Cosmology

7 AI in astrophysics overview

Speaker: Prof. Tobias Buck (University of Heidelberg)

8 Talk

Speaker: Alexandre Adam (University of Montreal)

9 Talk

Speaker: Dr Bruno Régaldo-Saint Blancard

16:00 Coffee break

16:00 Poster Session

Cosmology

10 Analysing edge-on galaxies with deep learning

The advent of large astronomical surveys, such as Euclid, will offer unprecedented insights into the statistical properties of galaxies. However, the large amounts of data that will be generated by these surveys call for the application of machine learning methods. For this purpose, we trained the YOLOv5 algorithm to accurately detect spiral, edge-on galaxies in astronomical images and the SCSS-Net neural network to generate segmentation masks, so that the detected galaxies can be used for any further analysis. This algorithm was applied on current astronomical images; however, its real power lies in its applicability to data from future surveys, where it can lead to new discoveries. We will also present one of our future goals, which is the study of the galactic warps: a well-known distortion of the galactic discs occurring in most spiral galaxies, including the Milky Way. Despite the fact that we know hundreds of warped galaxies of different shapes and sizes, it is still not clear how the warp is created. We will show how our algorithm can yield a deeper statistical analysis that will enable us to make connections between the different warps, understand their environmental dependencies, and thus contribute to understanding how this feature forms and what role it plays in the galactic evolution.

Speaker: Dr Žofia Chrobáková (Mullard Space Science Laboratory, University College London, Holmbury St Mary, Dorking, Surrey RH5 6NT, UK)

11 Emulating the Interstellar Medium Chemistry with Neural Operators

Understanding the interstellar medium (ISM) chemistry is pivotal for the study of galaxy formation and evolution. Traditional computational models rely on costly ordinary differential equation (ODE) solvers to simulate complex photo-chemical processes. This study introduces a novel approach using DeepONet neural operators to emulate a non-equilibrium chemical network, significantly reducing computational costs while maintaining precision. Unlike conventional methods, our approach approximates the differential operator directly, enabling the model to generalize beyond the specific conditions encountered during training. This capability ensures the robustness and flexibility of the emulation across a broader parameter space, including varied densities, temperatures, species abundances, and radiation fields. Remarkably, our method maintains an accuracy within 1\% while achieving speed-ups of up to 128× compared to traditional methods. This makes it a powerful tool for large-scale astrophysical simulations and advancing our understanding of ISM dynamics.

Speaker: Dr Lorenzo Branca (Interdisziplinares Zentrum fur¨ Wissenschaftliches Rechnen, Universitat¨ Heidelberg)

12 Neural Network Approaches for Quasar and Galaxy Continuum Estimation: A Comparative Study of Autoencoder and U-Net Architectures

The rise of new spectroscopic surveys, such as the WHT Enhanced Area Velocity Explorer (WEAVE), EUCLID, and the 4-metre Multi-Object Spectroscopic Telescope (4MOST), alongside the ongoing Dark Energy Spectroscopic Instrument (DESI), will significantly increase the volume of observed quasar and galaxy spectra. This surge necessitates the development of automated methods for accurate spectral continuum estimation. This study aims to evaluate the performance of two neural network (NN) architectures -— an autoencoder and a convolutional NN (CNN) known as U-Net -— in predicting quasar continua within the rest-frame wavelength range of 1020 Å to 2000 Å and galaxy continua within the range of 3500 Å to 5500 Å.

Speaker: Dr Francesco Pistis (Università degli studi Milano-Bicocca)

17:15 Discussion - AI in Astrophysics

Thursday, 26 September

Particle Physics

Conveners: Prof. Andrea Wulzer (ICREA and IFAE, Barcellona), Prof. Tilman Plehn (Heidelberg University)

13 Talk

Speaker: Prof. Andrea Wulzer (ICREA and IFAE, Barcellona)

14 Talk

Speaker: Prof. Tilman Plehn (Heidelberg University)

10:20 Coffee break

Particle Physics

Conveners: Mr Sam Bright-Thonney (Cornell Univeristy), Stefano Forte (Milan University)

15 Plenary talk

Speaker: Stefano Forte (Milan University)

16 Plenary talk

Speaker: Samuel Bright-Thonney (IAIFI Cambridge)

12:10 Lunch break

Econophysics

Conveners: Prof. Micheal Benzaquen (École Polytechnique & Capital Fund Management (CFM)), Riccardo Milocco (IMT Lucca)

17 Econophysics, why and what for?

Speaker: Prof. Michael Benzaquen (École Polytechnique & Capital Fund Management (CFM))

18 Multi-Scale Node Embedding and Network Reconstruction: A Comparative Analysis with the Single-Scale Literature

See Attached File

Speaker: Riccardo Milocco (IMT Lucca, ING Bank, Leiden University)

abstracts.pdf

19 Econophysics: talk

20 Econophysics: talk

16:00 Coffee break

Quantum Machine Learning

Conveners: Davide Pastorello (Università di Bologna), Prof. Giacomo De Palma, Sreetama Das (Università di Firenze)

21 Quantum machine learning for physics

Speaker: Prof. Giacomo De Palma

22 Quantum Neural Networks: An overview and some recent results

Speaker: Davide Pastorello (Università di Bologna)

23 Quantum Machine Learning: talk

Speaker: Sreetama Das (Università di Firenze)

Friday, 27 September

Model Complexity with AI

24 Plenary talk

Speaker: Prof. Fosca Giannotti

25 Conterfactuals Generation in Network Science, Econometrics and Explainable AI

Speaker: Marzio Di Vece (SNS)

26 Model Complexity with AI: talk

Speaker: Andrea Pugnana (Pisa University)

27 Causal inference over time

Speaker: Isacco Beretta (Pisa University)

11:00 Coffee break

Model Complexity with AI

28 Feature Learning and Generalization in Deep Networks with Orthogonal Weights

Speaker: Hannah Day (University of Illinois Urbana-Champaign)

29 AI and jobs: mapping forward-looking AI exposure metrics into occupational networks

To quantitatively investigate the AI impact on the world of work, the first challenge to address is how to measure AI’s impact on individual occupations.
A major issue with the existing approaches is that they mostly depend on expert evaluations of AI capabilities in relation to occupational tasks. This reliance makes the process of gathering these evaluations less transparent, objective, and reproducible. Moreover, they inherently measure the 'potential' impact of AI on occupations rather than its actual impact.
In this paper, we develop a novel occupational AI exposure index to measure the near-future actual exposure of occupations to AI, rather than their potential exposure. Using data on AI applications from venture capital-funded startups, our index assigns an exposure score to occupations by connecting (with a large language model) descriptions of AI applications developed by startups to job descriptions. Unlike existing indices, our measure effectively maps concrete future market directions.
We compare the occupational AI exposure scores generated by our new index with those from the widely-used AI Occupational Exposure (AIOE) index. To do this, we adopt a network perspective. In particular, we construct a network of jobs, where two jobs are connected if they require a similar set of abilities.
When using the AIOE index, we observe two big clusters in the job network, indicating that jobs with similar exposure levels cluster together. This implies the existence of a "potential AI trap", where workers attempting to move from an AI-exposed job to a job requiring similar skills, would likely end up in another job with similar potential exposure.
Conversely, when applying our new metric, we uncover a different scenario: many occupations considered potentially exposed to AI are not actually targeted by AI startup applications. This results in the emergence of more, smaller clusters. Consequently, our finding suggests that an actual AI trap is still far.

Speaker: Enrico Maria Fenoaltea (Centro ricerche Enrico Fermi)

AI_Pisa_workshop.pdf

30 Interpretable Neural-Symbolic Concept Reasoning

Speaker: Francesco Giannini (Scuola Normale Superiore)

12:30 Lunch break

Particle Physics

Conveners: Bryce Fore (Argonne National Lab), Dr Claudius G. Krause (OAW, Vienna), Mr Filippo Cattafesta (Pisa University and INFN Pisa)

31 Talk

Speaker: Claudius G. Krause (OAW, Vienna)

32 Modeling Low-Density Nuclear Matter with Neural-Network Quantum States

The structure of low-nuclear density nuclear matter is of great importance to the physics of neutron star crusts. One of the important aspects of this structure is the transition from roughly spherical neutron rich nuclei to uniform matter. Models for both of these extremes exist but the transition is less easily understood. In this presentation I will discuss my results from variational Monte Carlo calculations using neural-network quantum states which show that this method can model nuclear matter in this density region efficiently and more accurately than other Monte Carlo methods.
The results to be shown come from calculations at several densities and proton fractions using a pionless effective field theory Hamiltonian. From these results I will show predictions for clustering, symmetry energy, and proton fraction for the beta-equilibrated ground state.

Speaker: Bryce Fore (Argonne National Lab)

33 CMS FlashSim: end-to-end simulation with ML

Detailed event simulation at the LHC is taking a large fraction of computing budget. CMS developed an end-to-end ML based simulation that can speed up the time for production of analysis samples of several orders of magnitude with a limited loss of accuracy. As the CMS experiment is adopting a common analysis level format, the NANOAOD, for a larger number of analyses, such an event representation is used as the target of this ultra fast simulation that we call FlashSim. Generator level events, from PYTHIA or other generators, are directly translated into NANOAOD events at several hundred Hz rate with FlashSim. We show how training FlashSim on a limited number of full simulation events is sufficient to achieve very good accuracy on larger datasets for processes not seen at training time. Comparisons with full simulation samples in some simplified benchmark analysis are also shown. With this work, we aim at establishing a new paradigm for LHC collision simulation workflows in view of HL-LHC.

Speaker: Filippo Cattafesta (Scuola Normale Superiore & INFN Pisa (IT))

34 Talk

16:20 Closing aperitif

16:20 Poster Session