“So far, astronomers have dealt with images of the night sky. Now that we have film of the night sky, what we need are algorithms capable of taking this information and summarising it in a meaningful way.”
That was the plan according to astronomer Ioana Ciucă when she joined the School of Computing at The Australian National University (ANU) in April 2022 as a Jubilee Joint Fellow.
From a child stargazer in rural Romania to the global nexus of astronomy and advanced computing, her lifelong fascination with the origin of our galaxy has now been galvanised by a realisation that artificial intelligence (AI) can and must revolutionise the field.
As part of the fellowship, which is in partnership with ANU Research School of Astronomy and Astrophysics, Ciucă deployed Astro-Machine Learning (AstroML) to confirm the existence of a disc of stars near the heart of the Milky Way Galaxy that pre-existed a cosmic collision between our galaxy and a smaller one called Gaia-Enceladus 8 to 11 billion years ago.
She named this ancient structure after her grandmother, Babi.
“I thought it was important to have names that are not just English,” she said.
Pie in the sky
Ciucă’s research partner at ANU is Professor Yuan-Sen Ting, a world-leading expert in AstroML. Ting helped Ciucă apply Bayesian Neural Networks to a high-resolution spectroscopic survey of our galaxy, leading to the Babi disc confirmation.
In March of this year, Ciucă and Ting teamed with Kartheik Iyer, a NASA Hubble Fellow at Columbia University, to study other ways that Large Language Models (LLMs) can enhance scientific exploration. Their research, published just three months later, involves a globe-trotting, life-changing mad scramble as well as some startling scientific discoveries.
First, they fed their journal article about the Babi disc into OpenAI’s groundbreaking ChatGPT-4. Then, they jumped in front of a keyboard to quiz the LLM about their research. If the model failed to accurately interpret their work, Ciucă and Ting would know better than anyone.
“We found that it translates exceptionally well,” Ciucă said. “You can ask it, ‘Can you explain this in easier words?’ or ‘Can you translate this into Romanian’, and it does an excellent job.”
After uploading several more astronomical papers, Ciucă and Ting found that the LLM was capable of functioning as a research assistant that can locate and synthesise knowledge from a vast body of scientific literature.
They uploaded a thousand more papers and tested the model again.
“To our amazement, these Large Language Models, when given an extensive context, can come up with a robust and non-trivial scientific hypothesis that rivals the work of humans,” Ting said.
In their most recent iteration, Ting and Ciucă developed two bots to converse with each other. They instructed the first to serve as a hypothesis generator and the second to offer critique. They found that the more the two bots interacted, the better their theoretical reasoning became.
“It was so exciting,” said Ting. “Note, we did not train the model. We just stress-tested the existing models in an innovative way and checked the results with our domain expertise.”
When Ciucă and Ting published a preliminary research note in April, they sparked a debate on social media platforms frequented by astronomers. Detractors were not shy.
“There was a palpable tension rippling through the community, with early adopters promising great opportunities over the horizon, and sceptics rolling their eyes and expressing disdain,” Ting said.
Just a few weeks prior, Elon Musk and others had released an essay warning of the risks of AI. Ciucă and Ting’s trial balloon had lofted into a maelstrom of politicians, opinionators and luminaries predicting calamity and even human extinction at the hands of AI.
“People have a hard time adjusting to AI because they realise it can be quite disruptive,” Ciucă said. “But I think it’s better to learn about it and find ways to improve it and to use it for good, rather than staying on the sidelines saying AI is bad.”
As Ciucă engaged via social media with AI sceptics and supporters alike, she was buoyed when Alyssa Goodman, a renowned Professor of Applied Astronomy at Harvard University, weighed in.
It was only a “like” but the gesture of encouragement was a sign of big things to come.
Microsoft had reached out to Professor Goodman looking to fund research projects on using AI foundation models for science and the global good.
“She messaged me about partnering,” Ciucă said. “I couldn’t believe she was willing to lend such enormous resources to our pie-in-sky ideas!”
Two months later, Ciucă and Ting were jetting off to the United States, first to Redmond, Washington to meet with Microsoft and then to Cambridge, Massachusetts to kick off one of the pilot studies in the new funding scheme. Nicknamed UniverseTBD, the research team comprises 40 contributors spanning four continents and several universities, The European Space Agency (ESA) and the Smithsonian Astrophysical Observatory.
At Harvard, they are collaborating with the purveyors the Astrophysics Data System (ADS), an online database of over 16 million astronomy and physics papers developed by NASA.
“If ADS implements some of our ideas—for instance, we develop a new application attached to the search engine—it would be game-changing,” Ting said.
On 9 June, Professor Goodman tweeted: “Someday, future astronomers will know what ‘UniverseTBD’ meant to our field. Today, I was honored to see it taking off”.
AI will democratise astronomy
Ting and Ciucă met virtually in 2020 when Ting was in the United States and Ciucă was in the United Kingdom. Both were elated to have been offered appointments at ANU. Then, the pandemic hit, and Australia closed its borders.
Stranded, Ting and Ciucă began swapping ideas via video conferencing—a lifeline that helped them cope with months of isolation and laid the groundwork for future, in-person collaboration.
Astronomers at elite universities enjoy advantages that go beyond research funding and access to high-powered telescopes. Colleagues and mentors often meet on or near campus to provide feedback, guidance, and encouragement. They read each other’s work and practise presenting their papers.
By contrast, astronomers on less established university campuses lack colleagues with whom they can share knowledge.
Although LLMs cannot entirely replace the benefits of in-person collaboration, Ciuca and Ting believe they can level the playing field.
“This technology can connect us all to each other, in our own languages,” Ciucă said.
Ting was born in Malaysia and, like Ciucă, speaks English as a second language. English is the dominant language for astronomical research, and in many ways a barrier to entry for the field.
Ting and Ciucă are examples of astronomers who have made contributions despite that barrier, but they say there are thousands of astronomers and astrophysicists who could be doing more.
With the help of AI, that is likely to change.
“AI can help you to identify useful information from deep inside of the body of research,” Ting said. “We need humans to understand it, because computers are not scientists. But it’s very beneficial to have a partner or co-pilot, especially for those who face language barriers and other obstacles.”
Ciucă has high hopes for UniverseTBD, and faith that AI will open doors for untapped talent in astronomy and other scientific fields.
“When the internet first appeared, it disrupted business-as-usual and it took a while before people adapted,” Ciucă said. “It’s the same here. One concern is that the AI will replace scientists which is an unfounded fear. It relies on this idea of ‘us vs them’ rather than ‘us with them’, which is where the real opportunity lies. The best scenario is a co-pilot scenario.”
Ciucă said she recognises the ethical, economic, and political shockwaves AI has created. She expects it to create more.
“Let’s not fight for AI or against AI. If anything, let’s fight against the forces that want to turn AI against us,” Ciucă said.
Having glimpsed the impact that AI can have on astronomy, Ciucă and Ting are hopeful and passionate enough to alter the course of their academic careers.
“The stars outnumber astronomers 10,000,000,000,000,000,000 to one,” Ciucă said. “We need all the help we can get.”
———