Tatiana Acero-Cuellar

she/her/ella

tatiana
CV GitHub

Ph.D Candidate, University of Delaware
Unidel Distinguished Graduate Scholars Fellow

You can find publications I have contributed to here:

Google Scholar

Hi there!

I'm Tatiana Acero-Cuellar, I'm a Unidel Distinguished Graduate Scholar Fellow and a last-year Ph.D. Physics student at the University of Delaware.

I use machine learning to extract meaning from images; whether the instruments are pointing towards the Earth or outside, I'm there. I have direct experience on the Vera C. Rubin Observatory Alert Production commissioning team. I developed and implemented the Real-Bogus deep learning model for Rubin, assigning a reliability score to detected sources to distinguish astrophysical transients from artifacts. I led the Rubin Difference Detectives, a citizen science project on the Zooniverse platform. My research in astrophysics includes work on the development of novel methods for transient discovery and physics-driven simulations of light echoes in astrophysical images. Outside of astrophysics, I have led work on the automatic detection of fish ponds in Nigeria using satellite imagery. I hold a B.Sc in Physics from the Universidad Nacional de Colombia (National University of Colombia).

Finding ways to connect more with my culture, I was the president of the Hispanic/Latino Graduate Student Association (HLGSA) at the University of Delaware for the 2025-2026 period.

Research

Astrophysics

Deep Learning Models

cnn

I'm developing the Real-Bogus (a.k.a. the reliability) model for the Vera C. Rubin Observatory. I have already delivered three versions of the model, which are reported in the technical note [DMTN-337](https://dmtn-337.lsst.io/)(and a paper is coming up soon!). A version of the model was deployed to assign the reliability score for all detections in the [Data Preview 1](https://dp1.lsst.io/) data and [Data Preview 2 data](https://dp2.lsst.io/). The last and current version of the models is being deployed for the generation of the LSST-alert data. The machine learning reliability model for Rubin is a light Convolutional Neural Network (CNN) that assigns each detected source a reliability score in [0,1], where 1 indicates a real astrophysical source, and 0 indicates an artifact.

Before joining Rubin, I developed a Real-Bogus model using autoscan dataset (from the Dark Energy Survey), where I compared the performance of the classifier when training with the template, science, and difference image, and only with the template and the science.

See paper here

Model interpretability

Distinguising between real and bogus astrophysical objects can be more challenging than just a binary classification. I created an informative latent space of autoscan dataset using a contrastive learning technique called Boostrap Your Own Latent (BYOL). The plot is generated by reducing that latent space to a 2D space with UMAP. Similar things are closer in this space.

Light Echoes Simulation

le

Feed-Forward simulation of Light Echoes. I am generating a simulated data set of Light Echoes images using the physics principles behind their formation, and observed properties of the dust medium. Light Echoes are detected on difference images, I am using the LSST Science Pipelines to inject the Light Echo simulations and perfom Difference Image Analysis.

Remote Sensing

Deep Learning Segmentation for Fish ponds in Nigeria

le

I am part of an interdisciplinary project where I am applying the YOLOv7, a computer science segementation deep learning model, to find and segment fish ponds in Nigeria.

About Me

I was born and raised in Bogotá, Colombia. Proudly latina and rola 🇨🇴. I believe in the power of individual actions 🌎. I enjoy anything that makes me laugh and think. Self-proclaimed a powerlifter 💪.