TP-IV: EPFL Astrophysics offers students access to cutting-edge research in modern astrophysics and cosmology via practical work assignments, aiming to decipher existing observations, test established theories, and push forward innovative tools and methods to increase our understanding of the Universe.
The wide variety of these mini research projects reflects the broad range of topics and expertise covered by the different astrophysics research groups and laboratories at EPFL.
Applications from other disciplines are welcome – if you wish to participate in a project, please contact directly the indicated faculty member.
Note that some of the Master thesis projects can potentially be adapted to be TP-IVb. Please check both pages!
Proposed projects
Available as 2025-2026 TP-IV a & b (or other 8 credits projects).
Supervisor: Dr David Harvey / Ethan Tregidga
Recommended: This project is suitable for a student interested in machine learning, cosmology, dark matter, gravitational lensing.
Context: Dark matter makes up most of the matter in the universe, yet its nature remains one of the biggest open questions in physics. Machine learning now presents and exciting way to probe data in a new and efficient way and answer some of the biggest questions.
Project: In this project, you’ll build a self-supervised machine learning “foundation model” that learns from both cosmological simulations and real telescope data, with the goal of probing dark matter through galaxy clusters.
By capturing general features that work across multiple datasets, the model will be reusable for many other astronomy and cosmology problems, from detecting rare objects to classifying large-scale structures.
The project will involve deep learning, contrastive learning, and multimodal data, applying these methods to large datasets.
This research will contribute to the development of scalable flexible tools for next-generation astronomical surveys, enabling deeper insights into the nature of dark matter and large scale structure.
References
- https://academic.oup.com/mnras/article/531/4/4990/7697182
- https://academic.oup.com/mnras/article/527/3/7459/7452889
- https://arxiv.org/abs/2503.15312
- https://iopscience.iop.org/article/10.3847/2041-8213/abf2c7
- https://openreview.net/forum?id=w0iQslghXD
Supervisor: Dr David Harvey / Ethan Tregidga
Recommended: This project is suitable for a student interested in machine learning, cosmology, dark matter, gravitational lensing.
Context: Dark matter makes up most of the matter in the universe, yet its nature remains one of the biggest open questions in physics. Machine learning now presents and exciting way to probe data in a new and efficient way and answer some of the biggest questions.
Project: In this project, you’ll build a self-supervised machine learning “foundation model” that learns from both cosmological simulations and real telescope data, with the goal of probing dark matter through galaxy clusters.
By capturing general features that work across multiple datasets, the model will be reusable for many other astronomy and cosmology problems, from detecting rare objects to classifying large-scale structures.
The project will involve deep learning, contrastive learning, and multimodal data, applying these methods to large datasets.
This research will contribute to the development of scalable flexible tools for next-generation astronomical surveys, enabling deeper insights into the nature of dark matter and large scale structure.
References
- https://arxiv.org/abs/2210.02747
- https://arxiv.org/abs/2507.11842
- https://arxiv.org/abs/2307.08698
Proposed by: Elisabeth Rachith
Overview
The Lastro team is developing advanced tools to detect and analyze traces of space debris and satellites in archival data from large field of view telescopes.
When these objects cross the field of view, they leave characteristic streaks on astronomical images. By studying these traces, we aim to extract valuable information about the objects’ stability and rotation rates.
To achieve this, we have implemented a machine learning algorithm capable of detecting streaks in telescope images. Subsequent analysis involves techniques such as light curve extraction and period fitting to further characterize the detected objects.
Challenge
While the detection algorithm demonstrates strong performance, a significant number of false positives are observed during manual review. The main sources of these false detections include:
- Pixel bleeding from oversaturated stars
- Star diffraction spikes
- Filter fringe patterns
These artifacts can mimic the appearance of genuine space object streaks, complicating the analysis.
Objective
To address the issue of false positives, we propose the development of a dedicated classifier. This classifier will be trained to distinguish between true detections of space debris/satellites and common sources of false positives.
Project Tasks
The project will proceed through the following stages:
- Classifier Architecture Selection
- Identify and select a classifier architecture well-suited to the specific requirements of the problem.
- False Positive Class Review
- Systematically review the images produced by the detection algorithm.
- Categorize and document the different types of common false positives.
- Dataset Preparation and Annotation
- Select a representative set of images for training and validation.
- Annotate the images with the correct labels for supervised learning.
- Classifier Training
- Train the classifier using the prepared dataset.
- Fine-tune parameters to optimize classification performance.
- Performance Assessment
- Evaluate the classifier’s effectiveness using appropriate metrics (e.g., accuracy, precision, recall).
- Analyze and report on the classifier’s ability to reduce false positives without compromising true detections.
Evaluation
The outcomes of the project will be assessed through a written report and an oral presentation.
Supervisors: eSpace/LASTRO (Prof. Jean-Paul Kneib / Elisabeth Rachith)
Type of Project: Semester project (TP4b)
Duration: 14 weeks (Official start/end dates: September 8-December 19, 2025)
Submission of final report: January 8, 2026
Final Presentation: TBD
Recommended: This project is suitable for a student interested in astronomy and space sustainability.
Category: Space situational awareness
Proposed by: Pascale Jablonka
Precise knowledge of the evolution of galaxies depends very much on our capacity to determine the age and chemical abundance of their stellar populations. Thanks to their proximity, this is possible at the level of individual stars in the Local Group galaxies and the Milky Way, with the analysis of their spectra. New spectroscopic wide sky surveys are about to begin, in particular with the WEAVE spectrograph, dedicated to the detailed study of the halo of our Galaxy and its dwarf satellites.
Samples of hundreds of thousands of stars must be analyzed. These orders of magnitude make it impossible to apply conventional techniques, when samples had a few tenth of stars. They require the development of deep learning techniques, with the goal of obtaining the same high precision in the stellar chemical abundances.
The student will help improve some machine learning codes, as developed on model stellar spectra, and will test them on the first real observational datasets.
Proposed by: Pascale Jablonka
Space telescopes, such as HST or JWST, make it possible to distinguish all the details of galaxy morphologies, but only over small field of views. On the contrary, ground-based surveys are very deep and can cover the whole sky, but do not have the same image quality because of the variability in the observational conditions. We are developing a range of new image processing techniques, generically known under the name of deconvolution, that enable us to recover spatial resolution from the ground-based images. This project will be particularly relevant to the Euclid and LSST data, with ultimate goal to shed new light on how galaxies fuel their star forming activity and how and where morphological transformation occur.
Proposed by: Andrii Neronov, Ettore Zaffaroni, Ronald Scaria
Cosmic rays entering Earth atmosphere produce Extensive Air Showers (EAS) of high-energy particles that can can be observed by an array of particle detectors on the ground. This phenomenon has been discovered back in 1938 by Pierre Auger at Jungrfaujoch laboratory in Switzerland. The EAS detection technique is nowadays used to study the highest energy particles known in nature (ultra-high-energy cosmic rays reaching 100 Exa-electronvolt) and for observations of the highest energy gamma-rays (highest energy gamma-rays reach Peta-electronvolt). One possible type of simple particle detectors is a water tank equipped with a fast photosensor, a Photo-Multiplier Tube (PMT). High-energy particles passing through the tank with the speed faster than speed-of-light in water emit blue Cherenkov light that can be detected with PMT.
The goal of the work is to participate in development of a prototype Water Cherenkov Detector Array (WCDA) in Geneva Lake for observations of EAS with energies in the Tera-electronvolt to Peta-Electronvolt. This prototype will be used for a feasibility study of a large WCDA for measurement of cosmic ray electron spectrum up to highest energies and improvement of sensitivity of observations of gamma-ray sky in the Tera-electron to Peta-electronvolt energy range.
Proposed by: Andrii Neronov
Magnetic fields that are relic of epochs right after the Big Bang can still reside in the low density regions of the Large Scale Structure, between galaxies and galaxy clusters. Detection of these fields and measurement of their properties might provide a valuable “window” on physical processes that have operated in the Early Universe a fraction of a second after the Big Bang. Such measurement is possible with the methods of gamma-ray and radio astronomies, using new observational facilities: gamma-ray Cherenkov Telescope Array (CTA) and radio Square Kilometer Array (SKA). The gamma-ray measurement technique is based on the observation of extended glow around distant extragalactic sources produced by electromagnetic cascade developing along the gamma-ray beam during its propagation through the intergalactic medium. Radio technique is based on the observation of Faraday rotation of the polarised radio signal through the intergalactic medium.
The goal of the work is to get an overview of possible mechanisms of generation of magnetic fields during the first microsecond after the Big Bang, their evolution toward the present-day intergalactic magnetic field and to study a possibility to constrain the “cosmological magnetogenesis” scenario with CTA and SKA observations.
Proposed by: Andrii Neronov, Volodymyr Savchenko
The highest energy gamma-rays produced by sources like supernova remnants, pulsar wind nebulae, active galactic nuclei, interact in the Earth atmosphere and produce extensive air showers of high-energy particles Such showers appear as very fast “shooting stars” on the sky, if imaged at Giga-Hertz imaging rate with imaging atmospheric Cherenkov telescopes. A new system of telescopes based on this technique, Cherenkov Telescope Array Observatory (CTAO) is now in construciton and first Large Size Telescopes (LST) of CTAO are starting to take data.
The goal of the project is to get aquainted with the principle of operation of CTAO, methods of analysis of data, to analyze the first data of LST, for a selection of sources and to get an idea of the mechanisms involved in acceleration of high-energy particles and production of gamma-rays in these sources.
turbulence in the interstellar medium, and detect of nanohertz-frequency gravitational waves with pulsar timing arrays.
In this project, the student will work on developing a likelihood-analysis method for pulsar searches in the Fourier domain. By developing a parameterization for the expected pulsar signal in the Fourier domain, we can try to improve the sensitivity of Fourier-domain searches.
Radio astronomy provides a unique probe of astrophysical phenomena within and beyond our solar system. However, a new growing challenge of radio interferometers is Radio Frequency Interference (RFI). Some of the loudest sources of radio waves are not from astrophysical sources, but radio and TV broadcasts, high-speed wireless communications (e.g. cell phone networks and WiFi), and radar. Radio interferometers are often built in remote and radio-quiet locations to avoid these sources of RFI. However, transient radio sources in the sky such as satellites are much more difficult to avoid. As the number of active satellites has rapidly increased (from 1,000 in 2013 to over 5,000 in 2022), RFI has become a growing concern in the radio astronomy community[1].
Cosmology is entering a new era with the deployment of next-generation telescopes like eRosita, Euclid, CMB-S4, and SKA, which will conduct large multi-wavelength surveys. These instruments will collect an unprecedented amount of data, offering a unique opportunity to explore the natures of dark energy and dark matter (DM). However, fully leveraging this information requires direct comparisons with cosmological simulations—a task currently constrained by the limited volumes achievable with existing computational resources.
To address this challenge, machine learning-based emulators have emerged as a promising solution to accelerate the forward modeling of astrophysical observations. In this project, we will focus on generating maps of extragalactic gas properties (such as temperature and density) from underlying DM density fields, using conditional deep generative models, including pix2pix GANs (Isola et al., 2018), diffusion models (Ho et al., 2020), and stochastic interpolants (Albergo et al., 2023). Using the extensive CAMELS simulation suite (Villaescusa-Navarro et al., 2021), we aim to create realistic 2D projections of these fields and extend this work into 3D representations. Another critical aspect of this endeavor will be incorporating the effects of baryonic physics in the conditioning of the emulators.
Pulsating stars are extremely useful tools for astrophysics, since their light variations allow to measure distances and probe their interior structure. The current era of large time-domain surveys is revolutionizing our knowledge of pulsations and increases enormously their applicability for distance measurements. This allows us to unravel the structure of the Milky Way, the nearby Universe, and calibrate measurements of the expansion of the Universe.
In this project, we will use the all-sky survey TESS combined with the ESA space mission Gaia to obtain an unprecedented view of low-amplitude multi-periodic long-period variable stars in the Milky Way. The goal of the project will be to identify the variable stars through their variability, determine the variability properties, and calibrate the period-luminosity relations that allow to use them for distance measurements.
In this project, you will learn to
- access large astronomical data sets through online archives, such as the Gaia archive and MAST
- process large amounts of photometric time series using python
- use astrometric (positions, parallax, proper motion) and photometric data to maximum advantage
- distinguish different variability classes
- calibrate period-luminosity relations
… and more!
This project can be done as a TP-IVb, other 8 ECTS, or Master project.
Contact: Bastian Lengen, Richard Anderson
The VELOCE (VELOcities of CEpheids) project provides unprecedented time-series radial velocity data of 256 classical Cepheids, including 75 spectroscopic binaries. Modeling these data is challenging because several signals are present at once: large amplitude pulsations (speeds of 10-70 km/s over weeks), orbital motion (up to 25 km/s over years), and other modulations, such as multi-periodicity, fluctuating periods, time-variable amplitudes, etc.
The goal of this project will be to extend an existing Markov Chain Monte Carlo code that models orbital motion to include the signals due to pulsational variability. The project can be easily extended to treat fluctuating periods or amplitudes as well. The challenge will be to find efficient implementations that allow to infer a maximum of information from the VELOCE radial velocity curves without overfitting.
In this project, you will learn to
- work with optical spectra and high-precision radial velocity time series of pulsating stars
- develop MCMC analysis tools and sharpen your statistics skills
- contribute to a large Python code base (> 10000 lines) and an ongoing research program
… and more!
This project can be done as a TP-IVb, other 8 ECTS, or Master project.
Contact: Giordano Viviani, Richard Anderson
The VELOCE (VELOcities of CEpheids) project provides unprecedented time-series radial velocity data of 256 classical Cepheids, including 75 spectroscopic binaries. Modeling these data is challenging because several signals are present at once: large amplitude pulsations (speeds of 10-70 km/s over weeks), orbital motion (up to 25 km/s over years), and other modulations, such as multi-periodicity, fluctuating periods, time-variable amplitudes, etc.
The goal of this project will be to develop RV curve fitting methods using Regularization techniques to minimize the number of fit parameters used for representing pulsational variability. Regularization will improve the representation of Cepheid RV curves and allow to obtain more accurate orbital parameters, while also allowing the definition of RV template curves applicable to large spectroscopic surveys.
In this project, you will learn to
- work with high-precision radial velocity time series of pulsating stars
- develop regularization techniques for variability analyses and sharpen your statistics skills
- contribute to a large Python code base (> 10000 lines) and an ongoing research program
… and more!
This project can be done as a TP-IVb, other 8 ECTS, or Master project.
Contact: Giordano Viviani, Richard Anderson
Cepheids are pulsating stars whose radius and brightness vary within a stable period. This feature is particularly important in astrometry since it allows us to measure their distance accurately. And consequently, use them as standard candles to calibrate the cosmic distance ladder. However, many other effects other than their pulsation can modify the incoming signals from these stars, such as the presence of an orbiting star. In these cases, the spectra, and therefore the measured radial velocity, of the Cepheid will contain information from both phenomena that can be complicated to distinguish.
The aim of this project is to explore a newly developed methodology that could allow us to determine the pulsation and orbit periods of binary Cepheids without using any prior knowledge. This method constructs periodograms calculated using the concept of partial distance correlation, which allows us to effectively distinguish the Doppler shifts due to orbital motion and the spectral line variability induced by the stellar activity.
In this project, the student will work with part of the python package SPARTA and apply it to real study cases. The student will study the limitations and strong points of this method. Understand the precision and accuracy of the results. Propose modifications or improvements to the technique and experiment with them.
Links:
Method: https://ui.adsabs.harvard.edu/abs/2022A%26A…659A.189B/abstract
SPARTA: https://github.com/SPARTA-dev/SPARTA
This project can be done as a TP-IVb, other 8 ECTS, or Master project.
Contact: Giordano Viviani, Richard Anderson