“There is no position for an astronomy/AI person already formed.” Quite an admission from Kari Haworth, chief technology operator at the Harvard-Smithsonian Astrophysical Observatory. As the first engineer in an executive position there, she was speaking to an audience at SXSW in Austin. Faced with a deluge of data from ground-based and space-based observatories, the field of astronomy is faced with a major crisis that only AI can address. But the astronomers are clearly behind the curve on implementing AI.
The Webb Space Telescope is already generating lots of data. On the ground, “the Vera Rubin Observatory is going to see more in its first year than all other telescopes in history combined! That is an insane amount of data that we are about to get,” explained Haworth. “At the Center for Astrophysics (CfA), we’re joining the world in discovering what AI can do. This is a new golden age of astronomical discovery.”
CfA research scientist Philip Cargile talked about exoplanets: planets orbiting other stars. He posed the question, “How do you integrate AI into that type of research?” Thousands of such planets are now known, “but what we really want to do is characterization: a focused study on individual planets. We want to know things like composition, temperature and size. It requires a very dedicated technology to do this.”
The key goal of such research is the detection of biosignatures, which are indications of life based upon what we know about life. “The data we are getting is very information-rich, but it’s also very complex,” said Cargile. “To say something about we are seeing, we build models to try to match the atmospheres. If you think about how many free parameters you need to model the atmosphere of Earth, it is very complex. The traditional approach for this is someone sitting in front of a computer and fine-tuning these models to match the data. Based on the amount of data from James Webb and future missions, this is unattainable. We need better ways to do this inference work.”
The European Space Agency recently simulated a bunch of data for an upcoming space mission, and they presented the astronomical community with this question: can you design a model that can go directly from the data to the inferred parameters? One student (Mayeul Aubin) worked with a team of exoplanet modelling experts at the CfA, and also another team of people familiar with AI modelling. The model is a simulated-based inference so it’s using an AI tool called normalizing flow to go directly from the data to the inferred parameters: not only the best-fit parameters but a measure of the uncertainties and correlations between the parameters.” After Aubin trained the tool, his project won first place over 398 other teams, correctly inferring the parameters. It is a great example of what astronomers can do with AI. It’s easiest to think of AI as one organ of our future scientific missions, used for finding needles in haystacks.
Right now data is transmitted from spacecraft to Earth using radio frequencies. Laser communication from space observatories will allow data rates a thousand times higher than we have today. It will come as a surprise to most people, but a vast amount of data collected even by the most advanced observatory, the Webb space telescope, never reaches scientists on Earth. “This is really frustrating,” lamented Haworth. “I know what we’re throwing away: information about cosmic rays and the big bright things and the faint things. We can only bring down the minimum of what you need to find out what you want.” Again, once such a rapid data transmission becomes possible, only an AI-based system will be able to sort through it all. “The next wave of astronomical discovery is rooted in emerging technology,” said Juliana Cherston of CfA, as the astronomers present were keen on meeting AI people at SXSW to help them with that emerging tech.