Meet the AAS Keynote Speakers: Dr. Cecilia Garraffo

In this series of posts, we sit down with a few of the keynote speakers of the 244th AAS meeting to learn more about them and their research. You can see a full schedule of their talks here, and read our other interviews here!

Dr. Cecilia Garraffo is the “AI Astrophysicist” at the Harvard & Smithsonian Center for Astrophysics (CfA) and director of the AstroAI Institute there. Her work focuses on applying the newest techniques in machine learning and artificial intelligence to problems in astrophysics spanning from black hole binaries to protostellar disks. 

Originally from Argentina, Dr. Garraffo received her Master’s in Astronomy at the National University of La Plata and her Ph.D. in Physics from the University of Buenos Aires. Her Ph.D. was in theories of gravity, which led her to a position at Brandeis University, where she developed an interest in their neuroscience research group where she learned about neural networks. This interest launched her into the world of machine learning and artificial intelligence methods for astronomy. She was a postdoctoral fellow at the CfA doing computational astrophysics and then joined the Institute for Applied Computational Sciences at Harvard where she studied pure AI methods and taught data science courses. Dr. Garraffo then returned to the CfA to work on Chandra and in 2023 founded the AstroAI Institute, which she is now the full-time director of. 

Dr. Cecilia Garraffo (photo via the CfA)

Bridging the Astronomy-AI Gap

Dr. Garraffo’s research focuses primarily on developing and applying new AI methods for astronomy. She bundles AI and machine learning (ML) together under the AI label, as the two fields are really inseparable at this point. The tools in use include transformers, custom large language models (LLMs) fine-tuned on astrophysical literature, autoencoder neural networks, and generative AI diffusion models. 

Commercial applications of AI are not developed with the data and tasks of astronomy in mind. Astronomy data is very challenging: it’s sparse, biased, and comes in many different varieties, each with their own set of uncertainties. One of Dr. Garraffo’s projects is on detecting biomarkers (the signatures of molecules that might be indicators of life) in exoplanet atmospheres. By working with computer vision specialists at MIT – people who specialize in processing 2D data – they’re able to push the limits on their data by improving the signal to noise ratio (SNR) using correlations available in AI methods. 

The kinds of tasks astronomers want to do with data are also unique. Astronomers want to understand the universe, not reproduce it. LLMs and image generators are trying to produce things that look “real” to the human eye. Scientific models need to actually be real, with physics encoded in and no violations of physical laws. An image generator can put seven fingers on a hand, but a physics model can’t violate general relativity. Science “cannot afford to have hallucinations” – the model has to understand the domain it has been trained in, and cannot just be ‘a black box‘. We need models that we can trust, understand, and interpret. There’s also the challenge of the model needing to know enough physics to be accurate, but also allow us to explore new physics through them. 

These techniques are also applicable across fields with similar data and task challenges. Cecelia has recently begun a new program, EarthAI, doing similar method-focused work meant for earth science projects. It turns out that a codebase developed in AstroAI for x-ray binaries can also be applied in EarthAI’s estimation of methane abundances at different heights in the atmosphere. One of the founding ideas of the institutes is to pair together science domain experts and methods domain experts. It keeps the astronomers and earth scientists from being bogged down with the technicalities and the computer scientists from wandering off into aphysical models.

Advice You Wish You’d Gotten

Dr. Garraffo had lots of advice for students in the field. She empathizes so much with other international students in the field – “…it’s so much uncertainty and it is so hard to come into a system that you don’t know…you don’t know how the game works, nobody knows your supervisors, it’s really really hard so you have to hang in tight and not let the imposter syndrome get the best of you.” Trust in yourself and do it the best you can – it’s easy to fall into the trap of self doubt. 

It’s also okay (and even a good thing!) to be the person who knows the least in the room. The fear of not knowing enough can hold you back, but it doesn’t matter as long as you keep moving and learning. “Knowing who brings what to the table is something we have to try and appreciate.” You might not be the best statistician in the room, but you probably know a lot more about astronomy than the stats people do. Nobody is or needs to be an expert in everything.

Finally, be flexible in your career. “When anybody tells you their career path they’ll try hard to make it make sense, but it doesn’t always make sense…sometimes the planning path doesn’t work but it’s important to keep going if you want to keep going and to trust yourself and say okay, I can do this even if it takes some adaptability or flexibility.” The “successful” career path is not the only way or even the best way for many people. 

Dr. Garraffo’s plenary will cover why we need AI for astronomy to alleviate the challenges presented by our data and research questions, and some of the major techniques and problems that are being pursued.

To hear more about AI in Astronomy, tune into Cecilia Garraffo’s Plenary Lecture at 3:40pm CT on Tuesday June 11th at #AAS244! 

Edited by: Nathalie Korhonen Cuestas

Featured Image Credit: AAS

About Lindsey Gordon

Lindsey Gordon is a fourth year Ph.D. candidate at the University of Minnesota. She works on AGN jets, radio relics, MHD simulations, and how to use AI to study all those things better.

Discover more from astrobites

Subscribe to get the latest posts to your email.

Leave a Reply