Hi, I'm Mila

Welcome to my little corner of the universe

About Me

Hi there! I'm Mila, a physics grad with a focus on astrophysics and machine learning. I've been working as a research assistant in machine learning for about a year now. In my spare time, I've been learning front-end engineering, hoping to switch careers. I also like game development, run a Dungeons & Dragons campaign, crochet, and sometimes stream on Twitch.

Skills

Work Experience

Education

  • Data Science
    Python, Machine Learning (Pytorch, Tensorflow)
  • Web Development
    HTML, CSS and SCSS, JavaScript, React, Typescript, Tailwind
  • Game Development
    Godot, Unreal Engine 5
  • Version Control
    Git
  • Sep 2024 - Present
    Internship Software and Data Engineering, Cando Image
  • Aug 2023 - Jul 2024
    Research Assistant in Machine Learning for Complex Industrial Systems, ZHAW
  • Sep 2020 - Aug 2023
    Student Assistant, Faculty of Medicine UZH
  • Feb 2020 - Aug 2020
    Teaching Assistant in Physics for the Life Sciences, UZH
  • 2021 - 2023
    Masters in Physics, ETH
  • 2017 - 2021
    Bachelors in Physics, University of Zurich
  • 2013 - 2017
    Swiss Matura with focus on Business & Law, Kantonsschule Büelrain Winterthur
Download CV

Projects

PyCmap

Designed and implemented the Pycmap website using fast HTML, delivering a responsive and user-friendly platform now available online. Showcased skills in web development, design, and deployment.

Try out Pycmap

Flatland Escape

Designed and developed a minigame in Godot, including conceptualizing the idea, creating a development roadmap, implementing features, and delivering a fully functional game. Demonstrated skills in project planning, game design, and programming.

Try out Flatland Escape

Fermle - Daily Fermi Questions

A Fermi question is a problem that requires making rough estimates and logical reasoning to arrive at an approximate answer, often involving large quantities and limited data. Fermle will ask you one Fermi question a day upon which you get three tries to guess the answer.

Try out Fermle

Minigame A Month - Feb 2024 Game Jam

For the February 2024 "Minigame A Month" Game Jam with the theme "space" a friend and I built a decorating puzzle game called "Revamp My Space" using the game engine Godot. The goal of the game is to decorate a vampires lair to his liking. Along the way he will give you feedback and guidance on his preferences.

Try out Revamp My Space

3 Hour Game Jam: Trijam #255

For the 255th Trijam with the theme "darkness within" a friend and I built a small "runner" game called "Snake Snack". The player is lost at sea and gets swallowed by a sea serpent. The goal is to escape the serpent and collect as many sea shells along the way as possible.

Try out Snake Snack

Astrophysics

Master Thesis

Semester Project

Bachelor Thesis

The impact of Supermassive Black Holes on Galaxy Morphology

Abstract

This thesis deals with the impact supermassive black holes (SMBH) have on the evolution and morphology of galaxies. Three cosmological simulations were performed with the adaptive mesh refinement (AMR) hydrodynamical code RAMSES. Two of the simulations had SMBH’s implemented in the form of sink particles, whereas the third had none. Turbulence was resolved in the SMBH Bondi accretion process for one of the two sink particle simulations. For the analysis of the galaxies, a Sérsic profile was fit to the stellar and cold gas mass. Furthermore, the orbital circularity of the stellar mass was probed to further determine the structure of the galaxies. In addition, the SMBH behaviour of the sink particle simulations was evaluated. It was found that the galaxies with SMBH sink particles implemented were ellipticals whereas the non-SMBH galaxy was more of a spiral nature with an uncharacteristically large bulge. Adding turbulence to the accretion process seemed to delay the SMBH mass accretion and hence allows more time for star formation which leads to brighter galaxies.

Skills & Technologies

  • Python & Big Data: All the data analysis was done in Python. Since the analysed data consisted of very large simulation data stored in HDF5 files, the Python Package h5py was used to deal with it.
  • Fortran 90: Worked with the hydrodynamical code RAMSES which is written in Fortran 90.
  • Slurm: All the simulations were run on the cluster CSCS and thus I here first became familiar with the workload manager Slurm
Download Bachelor Thesis

Gallery

Mock-observation images of the simulated galaxies, once without (left) and with a SMBH (right):

Video of the time evolution of the formation of the simulated galaxy with SMBHs and turbulence, visualising gas density (left) and star density (right):

The Formation of Stellar Clusters at High Redshifts in Cosmological Simulations

Abstract

This report discusses the formation and evolution of globular clusters in three high redshift (z ≥ 6.0) galaxies. After identifying the globular clusters in the galaxies, their mass- and luminosity-functions are evaluated and discussed. Furthermore, the relation between the number of globular clusters and the halo mass is tested. The impact of the simulation mass-resolution as well as different feedback models is also examined. Finally, the formation sites of the the globular clusters are determined. It was found that the mass function of the clusters follows a power-law shape with a slope close to, but below, the theoretical value of β ≈ 2. On the other hand, the luminosity function followed a log-normal shape. Increasing the resolution of the simulation solely adds clusters at lower masses, without changing the shape of the mass function. Feedback, especially from supernovae, largely reduces the number of low-mass young clusters in a galaxy. Finally, it was found that stellar clusters form in regions of high density near the centre of the galaxy and later disperse.

Skills & Technologies

  • Python: All the data analysis was done in Python.
Download Semester Project

Gallery

A mock-observation image of galaxy B400-10, simulated with very high resolution (VHR), as well as its globular clusters are shown for redshift z=8 and z=6:

Video of time evolution of simulated galaxy B400-02 as well as its globular clusters and their age:

Convergence Maps With Denoising Diffusion Probabilistic Models

Abstract

In this thesis we investigate the efficacy of denoising diffusion probabilistic models (DDPM) in generating convergence maps conditional on the cosmological parameters Ωm and σ8, and compare it to previous work using the conditional Wasserstein-GAN (CWGAN) by Perraudin et al. 2020. We utilise a number of summary statistics, including the pixel value distribution, power-spectrum, bispectrum, Minkowski func- tionals as well as the Wasserstein-1 distance to compare the images generated by the DDPM to test data. Additionally, we examine the use of different hyper-parameters as well as sampling methods. We find that our DDPM successfully generates accurate convergence maps which are only marginally outperformed by the CWGAN and thus provide a simpler and more stable alternative. Given more time to fine-tune the DDPM, it is likely that it will be able to surpass the CWGAN.

Skills & Technologies

  • Python: All the data analysis was performed in Python.
  • Machine Learning: Denoising Diffusion Probabilistic Models, GANs, VAEs and Hyperparameter Optimization.
  • PyTorch: The DDPM was implemented using PyTorch.
Download Master Thesis

Gallery

Video of image generation process of the DDPM starting at step 2000 (gaussian noise) and ending at step 0 with the final convergence map:

Images of different cosmologies generated by the DDPM, with their position on the plot providing information on their cosmological parameters Ωm and σ8:

Publications

Data Scarcity in Fault Detection for Solar Tracking Systems: the Power of Physics-Informed Artificial Intelligence

Lüscher, M. F., Zgraggen, J., Guo, Y., Notaristefano, A., & Goren Huber, L. (2024). PHM Society European Conference, 8(1), 8.