Source: DeepTech
Astronomers are deploying large metal detectors in the deserts of Australia and South Africa. In the future, these detectors will work simultaneously to search for radio signals in the universe.
In the next five years, SKA (Square Kilometer Array Observatory) will be officially launched to find new information about the first star in the universe and the different stages of galaxy evolution.
But after hundreds of thousands of dish antennas and antennas are all synchronized, astronomers will soon face a new challenge. They need to sort out about 300PB of cosmic data every year, enough to fill one million laptops.
In the next decade, this problem will continue to arise at other observatories. As astronomers build more and more giant cameras to image the whole sky and send infrared telescopes to search for distant planets, the scale of data they collect will be unprecedented.
Cecilia Garraff, a computational astrophysicist at the Harvard Smithsonian Center for Astrophysics ） He said, "We are completely unprepared. Now we should be anxious. When you have too much data but no technology to process them, it is like no data. ”
In order to deal with the challenge of too much data, astronomers are turning their attention to artificial intelligence and optimization algorithms to capture hidden laws in large and complex data sets.
Some people are now working to establish research institutes that combine computer science and astronomy, and trying to build new partnerships.
In November 2022, Garaf launched AstroAI as a pilot project at the Harvard Smithsonian Center for Astrophysics.
Since then, she has formed an interdisciplinary team composed of more than 50 members and planned dozens of projects, focusing on the earliest appearance of the universe and whether we are the only life in it.
In the past few years, several similar cooperation organizations have followed the example of Mr. Galaf and are now competing for funds to expand to large institutions.
In his career in astronomy, physics and computer science, Garaf recognized the potential of artificial intelligence models. In the past cooperation, the language barrier was a major barrier she faced.
Usually, it is difficult for astronomers and computer scientists to work together because they use different words to describe similar concepts. Galaf is no stranger to translation challenges. She grew up in Argentina, but studied in a pure English teaching school and had similar experiences.
She tried to learn from this experience and tried to integrate people from the two communities so that they could identify common goals and find ways to communicate.
Astronomers have been using artificial intelligence models for many years, mainly to classify known objects such as supernovae in telescope data.
When the Vera C. Rubin Observatory is launched in 2025, this image recognition capability will become more and more important, because the number of supernova detections per year will rapidly jump from thousands to millions.
But the Sina wave of AI applications goes far beyond matching games. Recently, the algorithm has been optimized to perform "unsupervised clustering" ” In this clustering mode, they can capture the rules in the data without being told what to look for.
This opens the door for astronomers to find effects and relationships that they currently do not know. Gallaf said that these computing tools provide astronomers with the first "systematic search for the unknown ” Ability.
In January 2024, AstroAI researchers used this method to classify more than 14000 detection data from X-ray sources. Without this method, it would be difficult for them to finish the work.
Another way that AI works is to detect the chemical composition of the atmosphere of other planets. Astronomers use telescopes to analyze starlight that passes through the planet's atmosphere and is absorbed by different molecules at a specific wavelength.
To figure out the remaining spectrum, astronomers often compare it with the false spectrum they generate based on a few molecules of interest, such as water and carbon dioxide.
Exoplanet researchers hope to expand their search scope to cover hundreds of compounds that may indicate the existence of life on the star ball, but at present we need several weeks to find four or five compounds.
Thanks to the newly deployed James Weber Space Telescope and the Ariel Space Telescope scheduled to be launched by the European Space Agency in 2029, the number of exoplanets detected by us will increase from dozens to thousands, and the efficiency bottleneck of exploring compounds will become more and more troublesome.
Mercedes L ó pez Morales, an astronomer who studies the atmosphere of exoplanets at the American Astrophysical Center, said that processing all these observations "was originally an impossible task for us to complete. Things like AstroAI appeared just in time, just before the torrent of data hit us. ”
In 2023, Lopez Morales cooperated with Mayeul Aubin, an undergraduate student who was interning at AstroAI at that time, to establish a machine learning model that can more effectively extract molecular composition from spectral data.
Within two months, their team built a model that can search the spectrum of thousands of exoplanets in 31 seconds, looking for the characteristics of five different molecules. This achievement won them the highest award of the European Space Agency Ariel Data Challenge.
The researchers hope to train a model to find hundreds of extra molecules, so as to improve their chances of finding signs of life on distant planets.
AstroAI's collaboration has also produced a real simulation of black holes and maps of the distribution of dark matter throughout the universe.
Gallaf's goal is to eventually build a large language model similar to ChatGPT. The model is trained based on astronomical data, can answer questions about observation, and analyze literature to obtain supporting evidence.
Daniela Huppenkothen, an astronomer and data scientist at the Netherlands Space Research Institute, said: "It's like there is a huge new playground to explore. We can use artificial intelligence to solve problems that could not be solved before, because the calculation cost of these problems is too high. ”
However, as outlined by Hu Pengkesen in his recent preprint, there are also trade-offs in integrating AI into the astronomical workflow. Although AI models are efficient, they often run in ways that scientists cannot fully understand.
This opacity complicates their debugging and makes it difficult to determine how they introduce bias. Like all generative AI, these models are prone to hallucinations and non-existent relationships, and finally report their conclusions in an unwarranted confident tone.
Hupenkerson said: "It is important to critically examine the role and failure of these models. Otherwise, we will see how the universe works from a perspective contrary to the facts. ”
Researchers are trying to incorporate the error into the algorithm response to explain the new uncertainty. Some believe that these tools can ensure an additional level of review in the current publication and peer review process.
"As human beings, we naturally tend to believe in machines. ” Viviana Acquaviva, an astrophysicist and data scientist at the City University of New York, said that she recently published a textbook on the application of machine learning in astronomy.
"We need to be very clear about the results that are often not clearly explained, and very honest about our ability. ”
Researchers are aware of the ethical consequences of introducing AI, even in seemingly harmless environments such as astronomy.
For example, if only specific institutions can obtain computing resources to run these tools, these new AI tools may perpetuate existing inequalities in this field.
Hupenkerson said that if astronomers reuse existing AI models trained by the company for other purposes, they will "inherit many moral and environmental problems inherent in these models ” 。
Gallaf is trying to solve these problems. AstroAI models are open source and free. The organization will also help to adapt the model to different astronomical applications.
She also cooperated with the Berkman Klein Internet and Society Center of Harvard University to formally train the AI ethics team and learn best practices to avoid prejudice.
Scientists are still studying how the advent of artificial intelligence may affect astronomy. If AI models can propose new ideas and provide scientists with new research approaches, it will forever change the role of astronomers in deciphering the universe.
But even if AI is still only an optimization tool, it will also become a sharp weapon in the weapons arsenal of space exploration.
"AI will change the rules of the game. ” Gallaf said, "We can't just rely on ourselves. ”
Support: Ren
Typesetting: Liu Yakun
