A new emulator code for Cosmology - an application of machine learning algorithms to speed up cosmological analyses

Aug 4, 2026

Upload by Patrick

A machine learning-based emulator code accelerates cosmological analyses by reproducing theoretical models in very short times with high accuracy. The work integrates an emulator into CosmoBolognaLib, addressing the runtime bottleneck of Bayesian inference where cosmological functions must be evaluated millions of times. Using CosmoPower, a neural network emulates the two-point correlation function across varied cosmological parameters, with validation against the original model over the full input range. Bayesian parameter inference on DUSTGRAIN-pathfinder simulations shows near-perfect agreement while reducing execution time from tens of hours to seconds.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages94

Reading Time3 h 57 min

AccessFree to download

Abstract

Motivation: machine learning in cosmology

Emulator concept and integration into CosmoBolognaLib

Neural network emulator using CosmoPower

Training, validation, and parameter space coverage

Application to cosmological simulations and Bayesian inference

Scuola di Scienze Dipartimento di Fisica e Astronomia "Augusto Righi"Corso di Laurea Magistrale in Astrofisica e Cosmologia A new emulator code for Cosmology: an application of machine learning algorithms to speed up cosmological analyses Relatore: Prof. Federico Marulli Correlatrice: Dott.ssa Sofia Contarini Presentata da: Pierpaolo Nicolosi Anno Accademico 2021/2022 Abstract In recent years, artificial intelligence has become a part of our daily life. Machine learning algorithms are now used in many sectors, from telephony to entertainment platforms. In the same way, the field of Cosmology has benefited from the enormous advantages offered by this technology. More and more scientific researchers apply nowadays machine learning algorithms to process and analyze the huge amount of data provided by wide-field surveys and cosmological simulations. The number of publications related to the usage of these techniques in the astrophysical sector has indeed grown exponentially in recent decades, driven by the enormous successes that have been achieved. Our Thesis work fits perfectly into this context. What we present is a machine learningbased code aimed at drastically speeding up cosmological analyses. It is in fact an emulator, i.e. an algorithm able to reproduce a given theoretical model in a very short time and with great accuracy. We implemented the code into the public libraries CosmoBolognaLib (Marulli, Veropalumbo & Moresco, 2016), providing the users with a powerful tool to emulate the cosmological functions already present in these libraries. The main limitation of these functions is their run time in the context of Bayesian analyses. Indeed, having to calculate these models millions of times, this type of statistical analysis becomes extremely slow. Exploiting the machine learning algorithms provided by the numerical library CosmoPower (Spurio Mancini et al. , 2022), we built a neural network aimed at imitating the output of the theoretical model of the two-point correlation function, a type of statistic widely used in Cosmology. As a first implementation example, we focused on emulating the model by varying four cosmological parameters: the total matter density parameter, Ωm, the energy density parameter, Ωde, the amplitude of the primordial power spectrum, As and the redshift, z. The training process of the network was carried out using the large computational resources provided by the Department of Physics and Astronomy of the University of Bologna. We then validated the emulator by comparing its output to that of the original function, verifying its accuracy for the full range of input parameter combinations. Finally, we applied our code to the analysis of cosmological simulations. In particular, we measured the two-point correlation function of dark matter particles at three different redshifts of the DUSTGRAIN-pathfinder (Giocoli, Baldi & Moscardini, 2018 ; Hagstotz et al. , 2019) . We performed a Bayesian analysis to derive the posterior probability distribution of the parameters Ωm , Ωde and As, using both the original function of the CosmoBolognaLib and its emulated version. The results obtained with our emulator show almost perfect correspondence with that of the original model, with small differences that completely fall within the statistical fluctuations of the method. The really innovative part, however, lies in the timing of the analysis. With our emulator, the code becomes thousands of times faster, bringing the total execution time from several tens of hours toa few seconds. Our code will also be improved by extending the emulation to other cosmological functions and expanding the emulator’s range of validity to cover a wider parameter space. The potential applications of this methodology in the future are numerous, and its use will soon become a key element for future cosmological analyses conducted within the CosmoBolognaLib. Sommario In questi ultimi anni, l’intel

FAQs

What problem does the emulator code address in cosmological analyses?

It targets the long runtime of cosmological function evaluations in Bayesian analyses, which require millions of model calculations.

What is the emulator in this thesis?

An emulator is an algorithm that reproduces a theoretical model rapidly and accurately, enabling faster cosmological computations.

How is the emulator validated and where is it applied?

The neural network output is compared to the original two-point correlation function across the full input parameter range, and it is then used to perform Bayesian inference on DUSTGRAIN-pathfinder simulation data at multiple redshifts.

Aug 4, 2026

Upload by Patrick