A new machine-learning framework to generate star cluster models - Master Thesis in Physics of Data

Aug 3, 2026

The birthplaces of stars are turbulent regions where collapsing interstellar gas fragments into star-forming cores, producing complex substructure. Hydrodynamical simulations can model the formation process but demand substantial computational resources, and embedded primordial clusters are observationally difficult to constrain. This thesis introduces machine-learning tools trained on a limited set of hydrodynamical simulations to generate realistic initial conditions. A differentiable Gaussian process framework enables seamless integration into downstream inference pipelines, using a two-step training and sampling procedure. Multiple sampling strategies are compared to identify samplers yielding realistic stellar cluster realizations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages56

Reading Time2 h 21 min

AccessFree to download

Abstract

Introduction

Modeling star clusters with Gaussian Processes

Hydrodynamical simulations of star clusters

Generating artificial clusters

Methods

Gaussian Processes

Density Estimator

Sampling Methods

Energy based Markov Chain Monte-Carlo

Introduction

Incorporated physics

Algorithm description

Results

Processing pipeline

Sampling from 7D model

Energy based sampling

Conclusion

University of Padova Department of Physics and Astronomy “Galileo Galilei“Master Thesis in Physics of Data A new machine-learning framework to generate star cluster models Internal Supervisor Master Candidate Dr. Giuliano Iorio George-Pantelimon Prodan University of Padova External Supervisors Student ID Dr. Mario Pasquato 2046802 CIELA institute, Montréal, Canada Prof. Michela Mapelli Heidelberg University, Germany Academic Year 2022-2023 ii iv Abstract The birthplaces of stars are complex places, where turbulent interstellar gas collapses and fragments into star-forming cores, giving rise to non-trivial substructure. While the formation process can be modelled with hydrodynamical simulations, these are quite expensive in terms of computational resources. Moreover, primordial star clusters that are still embedded in their parent gas cloud are hard to constrain observationally. In this context, most efforts aimed at simulating the dynamical evolution of star clusters assume simplified initial conditions, such as truncated Maxwellian models. We aim to improve on this state-of-the-art by introducing a set of tools to generate realistic initial conditions for star clusters by training an appropriate class of machine learning modelson a limited set of hydrodynamical simulations. In particular, we will exploit a new approach based on Gaussian process (GP) models, which have the advantage of differentiability andofbeing more tractable, allowing for seamless inclusion in a downstream machine learning pipeline e.g. for inference purposes. The proposed learning framework is a two-step process includingthe model training and the sampling of new stellar clusters based on the inference results. We investigate different sampling approaches in order to find samplers that are able to generate realistic realizations. vi Contents Abstract v List of figures ix List of tables xi Listing of acronyms xiii 1 Introduction 1 1.1 Modeling star clusters with Gaussian Processes ................ 2 1.2 Hydrodynamical simulations of star clusters ................. 3 1.3 Generating artificial clusters .......................... 4 2 Methods 7 2.1 Gaussian Processes .............................. 7 2.2 Density Estimator ............................... 9 2.3 Sampling Methods .............................. 10 2.3.1 Markov Chain Monte-Carlo ..................... 10 2.3.2 Approximate Posterior Ensemble Sampler .............. 11 2.3.3 Rejection Sampling .......................... 13 3 Energy based Markov Chain Monte-Carlo 15 3.1 Introduction ................................. 15 3.2 Incorporated physics ............................. 16 3.3 Algorithm description ............................. 17 3.3.1 Define an energy space ........................ 17 3.3.2 Gaussian Processes modeling ..................... 18 3.3.3 Sampling procedure ......................... 18 4 Results 21 4.1 Processing pipeline .............................. 21 4.1.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 4.1.2 Models and training ......................... 21 4.1.3 Data pre-processing .......................... 23 4.1.4 Sampling and evaluation ....................... 26 4.2 Sampling from 7D model ........................... 27 4.2.1 Rejection sampling .......................... 27 4.2.2 Metropolis algorithm ......................... 27 4.2.3 APES algorithm ........................... 29 4.3 Energy based sampling ............................ 31 4.3.1 Hyper-parameters tuning ....................... 31 4.3.2 Newly generated clusters ....................... 32 4.3.3 Comparison to other methods .................... 33 5 Conclusion 37 References 39 Listing of figures 1.1 Open star cluster images ........................... 3 4.1 Distributions of inter-particle distances (a), velocity (b), and mass (c) for three simulated clusters: m1e4, m4e4, and m9

FAQs

Why are hydrodynamical simulations for star cluster formation challenging?

They are computationally expensive. Additionally, primordial star clusters embedded in their parent gas clouds are difficult to constrain observationally.

What machine-learning approach does the thesis propose?

It introduces a framework trained on a limited set of hydrodynamical simulations, leveraging differentiable Gaussian process (GP) models to generate realistic initial conditions.

How are new star clusters generated and evaluated in the framework?

The workflow is two-step: train the model, then sample new clusters using inference results. Different sampling strategies are investigated to find samplers that produce realistic realizations.

Aug 3, 2026

Upload by Lucas Martin