Master's Thesis Defense by Alejandro Maza Villalpando

Title: Machine Learning Methods for Predicting Stellar Parameters in Realistic Molecular Cloud Environments

Abstract: This thesis introduces a novel methodology that utilizes data from the RAMSES and MESA frameworks to employ machine learning techniques. The aim is to predict essential stellar parameters, namely temperature, stellar luminosity, radius, and accretion luminosity. The predictions used as inputs are based on values of mass, age, and accretion rate. These input values can be obtained readily in models of a realistic molecular cloud environment. The stellar structure models from MESA diverge from traditional methods that solely rely on mass and age. The investigation resulted in the implementation of several machine learning models. Using the results from these models, an extensive examination of their strengths and weaknesses was carried out. Furthermore, a comprehensive comparative analysis was conducted, directly contrasting these models with traditional methods like those presented in the literature by DM97 [1]. Based on this comparison, it can be concluded that machine learning methods exhibit high effectiveness as inference algorithms for predicting stellar structure parameters, especially in a dynamic accretion scenario, and exceed the performance of the DM97 model in this context. Aninference module was developed for Python and Fortran, enabling easy integration into simulation frameworks like RAMSES. Additionally, comprehensive documentation was also created to facilitate the incorporation of the models from this thesis into any other programming language. Finally, this thesis investigates potential improvements for the obtained models, outlining different perspectives such as enhancing data quality and investigating alternative machine learning architectures.

Supervisor: Troels Haugbølle

Censor: Maximilian Stritzinger, Aarhus University