Projects

One of the biggest challenges in cosmology and astrophysics is to understand the parameters that can describe the content, geometry and evolution of our Universe. The gravitational lens effect, produced by the deflection of the light rays by a gravitational field, is one of the most powerful tool to address this problems. There is a broad variety of scientific questions that can be studied using this effect, including the study of dark matter and dark energy. We are looking for more lens gravitational systems to address this questions with robust evidence.

I’m presenting here some of the research topics of my interest including the search for new gravitational lens systems.

If you want to see my list of publications please click here

I just create a repository for my scientific presentations so you can find the pdfs here

back

Autoencoder for deblending.

==Work In Progress==

Lens systems are such a wonderful tool, but they can be a huge headache when we can not disentangle how much light is coming from this or that object. Some arcs or counter images are usually blended with the lens galaxy, the solution so far have been modeling the system to try to deblend, but this is SOOOO time consuming!

So we have been creating many many simulations to train CNN for classification so why don’t we use them to train an autoencoder to deblend our systems? well after 10 failed tries… the 11 version of my autoencoder looks like it is doing something interesting ;)

But here you have a teaser:

Who classifies gravitational lenses better: professors, postdocs, students or the machine?

In this project my goal was to test how good humans are at classifying gravitational lensing and detect how we are biasing the selection of this system, as we have been always relying on human experts’ final word after a machine learning classification.

We recruited a total of 55 people that completed more than 25% of the project. During the classification task, we present to the participants 1489 images. We designed the data sets to have a broad variety of examples including: simulated lenses; real lens candidates; non-lens examples; and unlabeled data. We also duplicated 105 cutouts to investigate the level of consistency of individuals when classifying. We used the citizen science web portal Zoouniverse to host our project.

If you want a know who perform better professors? students? the machine? confident people? those with more than 10 years of experience in the field? you can read the full article here.

But here you have a teaser:

Convolutional Neural Network

We trained an Efficienet B1 with the goal to find new gravitational lens system in the Dark Energy Survey. A 0.4% of the whole sample analyzed by the convolutional neural network obtained a score above 0.9, which were then visually inspected. We found a total of 186 candidates that are newly identified by our search and another 219 previously identified for other works and methods.

In this project we also used Muscadet for deblending and we modeled the best candidates with only one deflector, using an automatic pipeline developed by a master student.

All the results can be found in here.

Strong lensing simulations

To use Convolutional Neural Network (CNN) to look for lenses in ground base data we require training/testing/validation sets of thousands of systems to be able to learn how they “look”. But there are only hundreds of known strong lensing systems, so they are not enough to be used in these sets, therefore the first and very challenging step is to simulate gravitational lens systems.

I developed a full data driven lens simulation toolbox based on Lenstronomy (Birrer & Amara 2018) to create realistic simulations based on real data unlike most of the previous work whose training sets were fully simulated or partially simulated.

back

Quasar Microlensing

Is produced when the images of the deflected source (Quasar) have about micro-arcseconds of separations, this effect can be produced by stellar objects, compact object (e.g. black holes) or stars, in the halo of lens galaxies. In this topic I have been work in two applications:

Study of the inner structures of QSOs.

This study is a big challenge, because this region is small and it is not resolvable with current observational facilities, even with VLTI. As a consequence, theories of accretion disks, which predict accretion disk sizes and their radial temperature profile (Shakura 1973), still need to be proven. This can be studied with indirect observational evidence, such as QSO microlensing that is produced by stars in the lens galaxy halo.

Using the microlensing properties I analyzed spectroscopic data searching for chromatic microlensing effect due to the source size being color-dependent. This allow us to estimate the size and the temperature profile of the accretion disk of quasars.

Hubble Constant and microlensing.

Time delay in strong lensed quasars offers an independent measurement of the expansion rate of the Universe, hence, the Hubble Constant. The COSMOGRAIL collaboration is aimed to measure high-precision time delays of known lensed quasars, using optical light curves in 6-7 month of monitoring. The final goal of COSMOGRAL, together with H0LICOW collaboration, is to provide a precise value for the Hubble constant.

In this context I investigate the possible systematic errors produced by microlensing in the quasar light curves as Tie & Kochanek (2018a) proposed. I developed a method to reproduce their work and investigate how mitigate this effect of microlensing time delay in COSMOGRAIL time delays measurements. This method is now the standard method in H0LiCOW to evaluate the impact of microlensing time delay on any new system measured by the collaboration.

back