SDSS J1038+4849 — image credit: NASA & ESA.
One of the biggest challenges in cosmology and astrophysics is to understand the parameters that describe the content, geometry and evolution of our Universe. The gravitational lens effect — the deflection of light rays by a gravitational field — is one of the most powerful tools to address these problems. It allows the study of dark matter, dark energy and other fundamental components of the Universe.
Here are some of the research topics I work on, including the search for new gravitational lens systems using artificial intelligence.
Lens finding
I use convolutional neural networks (CNNs) to identify gravitational lenses in ground-based imaging data. This machine learning approach requires large training, validation and test sets made of thousands of images. However, with only a few hundred known strong lens systems available, synthetic data is necessary.
To address this, I developed a simulation pipeline using Lenstronomy (Birrer & Amara 2018) to generate realistic simulations based on real observational data, unlike previous approaches relying solely on synthetic inputs.
The next steps include follow-up campaigns with powerful telescopes to confirm our CNN-selected candidates through high-resolution imaging and spectroscopy.
Humans vs. AI in lens classification
Every lens search ends with people looking at images, yet the reliability of this step is rarely measured. I led an experiment in which experts at different career stages — professors, postdocs, PhD and Master students — and citizen scientists classified the same set of images, and compared their performance with neural networks.
The results, published in MNRAS (2023), quantify the purity and completeness of human classification and inform how future surveys should combine visual inspection with machine learning.
Quasar microlensing
Quasar microlensing occurs when the light from a distant quasar is magnified by stars or compact objects in a foreground galaxy, producing micro-arcsecond scale variations. I have worked on two applications:
Inner structure of quasars
The central regions of quasars, including their accretion disks, are too small to be directly imaged, even with interferometers like the VLTI. Microlensing offers indirect access to this regime.
By analysing spectroscopic data for chromatic microlensing (wavelength-dependent magnification), I have estimated the size and temperature profile of accretion disks, testing models such as Shakura & Sunyaev (1973).
Hubble constant and microlensing
Strongly lensed quasars provide an independent measurement of the Hubble constant (H0) through time-delay cosmography. Projects like COSMOGRAIL and H0LiCOW use long-term photometric monitoring to measure these delays.
Microlensing-induced variations can bias these measurements. I studied and implemented the method proposed by Tie & Kochanek (2018) to simulate and correct for the microlensing time delay, making H0 estimates more robust. This methodology is now part of the H0LiCOW collaboration’s workflow.