Imagine if you could look at a snowflake at the South Pole and determine the size and the climate of all of Antarctica. Or study a randomly selected tree in the Amazon rain forest and, from that one tree—be it rare or common, narrow or wide, young or old—deduce characteristics of the forest as a whole. Or, what if, by looking at one galaxy among the hundred billion or so in the observable universe, one could say something substantial about the universe as a whole? A recent paper, whose lead authors include a cosmologist, a galaxy-formation expert, and an undergraduate named Jupiter (who did the initial work), suggests that this may be the case. The result at first seemed “crazy” to the paper’s authors. Now, having discussed their work with other astrophysicists and done various “sanity checks,” trying to find errors in their methods, the results are beginning to seem pretty clear. Francisco Villaescusa-Navarro, one of the lead authors of the work, said, “It does look like galaxies somehow retain a memory of the entire universe.”
The research began as a sort of homework exercise. Jupiter Ding, while a freshman at Princeton, wrote to the department of astrophysics, hoping to get involved in research. He mentioned that he had some experience with machine learning, a form of artificial intelligence that is adept at picking out patterns in very large data sets. Villaescusa-Navarro, an astrophysicist focussed on cosmology, had an idea for what the student might work on. Villaescusa-Navarro had long wanted to look into whether machine learning could be used to help find relationships between galaxies and the universe. “I was thinking, What if you could look at only a thousand galaxies and from that learn properties about the entire universe? I wondered, What is the smallest number we could look at? What if you looked at only one hundred? I thought, O.K., we’ll start with one galaxy.”
He had no expectation that one galaxy would provide much. But he thought that it would be a good way for Ding to practice using machine learning on a database known as CAMELS (Cosmology and Astrophysics with MachinE Learning Simulations). Shy Genel, an astrophysicist focussed on galaxy formation, who is another lead author on the paper, explained CAMELS this way: “We start with a description of reality shortly after the Big Bang. At that point, the universe is mostly hydrogen gas, and some helium and dark matter. And then, using what we know of the laws of physics, our best guess, we then run the cosmic history for roughly fourteen billion years.” Cosmological simulations have been around for about forty years, but they are increasingly sophisticated—and fast. CAMELS contains some four thousand simulated universes. Working with simulated universes, as opposed to our own, lets researchers ask questions that the gaps in our observational data preclude us from answering. They also let researchers play with different parameters, like the proportions of dark matter and hydrogen gas, to test their impact.
Ding did the work on CAMELS from his dorm room, on his laptop. He wrote programs to work with the CAMELS data, then sent them to one of the university’s computing clusters, a collection of computers with far more power than his MacBook Air. That computing cluster contained the CAMELS data. Ding’s model trained itself by taking a set of simulated universes and looking at the galaxies within them. Once trained, the model would then be shown a sample galaxy and asked to predict features of the universe from which it was sampled.
Ding is very humble about his contribution to the research, but he knows far more about astrophysics than even an exceptional first-year student typically does. Ding, a middle child with two sisters, grew up in State College, Pennsylvania. In high school, he took a series of college-level astronomy courses at Penn State and worked on a couple of research projects that involved machine learning. “My dad was really interested in astronomy as a high schooler,” Ding told me. “He went another direction, though.” His father is a professor of marketing at Penn State’s business school.
Artificial intelligence is an umbrella concept for various disciplines, including machine learning. A famous early machine-learning task was to get a computer to recognize an image of a cat. This is something that a human can do easily, but, for a computer, there are no simple parameters that define the visual concept of a cat. Machine learning is now used for detecting patterns or relationships that are nearly impossible for humans to see, in part because the data is often in many dimensions. The programmer remains the captain, telling the computer what to learn, and deciding what input it’s trained on. But the computer adapts, iteratively, as it learns, and in that way becomes the author of its own algorithms. It was machine learning, for example, that discovered, through analyzing language patterns, the alleged main authors of the posts by “Q” (the supposed high-ranking government official who sparked the QAnon conspiracy theory). It was also able to identify which of Q’s posts appeared to be written by Paul Furber, a South African software developer, and by Ron Watkins, the son of the former owner of 8chan. Machine-learning programs have also been applied in health care, using data to predict which patients are most at risk of falling. Compared with the intuition of doctors, the machine-learning-based assessments reduced falls by about forty per cent, an enormous margin of improvement for a medical intervention.
Machine learning has catapulted astrophysics research forward, too. Villaescusa-Navarro said, “As a community, we have been dealing with super-hard problems for many, many years. Problems that the smartest people in the field have been working on for decades. And from one day to the next, these problems are getting solved with machine learning.” Even generating a single simulated universe used to take a very long time. You gave a computer some initial conditions and then had to wait while it worked out what those conditions would produce some fourteen billion years down the line. It took less than fourteen billion years, of course, but there was no way to build up a large database of simulated universes in a timely way. Machine-learning advances have sped up these simulations, making a project like CAMELS possible. An even more ambitious project, Learning the Universe, will use machine learning to create simulated universes millions of times faster than CAMELS can; it will then use what’s called simulation-based inference—along with real observational data from telescopes—to determine which starting parameters lead to a universe that most closely resembles our own.
Ding told me that one of the reasons he chose astronomy has been the proximity he feels to breakthroughs in the field, even as an undergraduate. “For example, I’m in a cosmology class right now, and when my professor talks about dark matter, she talks about it as something ‘a good friend of mine, Vera Rubin, put on the map,’ ” he said. “And dark energy was discovered by a team at Harvard about twenty years ago, and I did a summer program there. So here I am, learning about this stuff pretty much in the places where these things were happening.” Ding’s research produced something profoundly unexpected. His model used a single galaxy in a simulated universe to pretty accurately say something about that universe. The specific characteristic it was able to predict is called Omega matter, which relates to the density of a universe. Its value was accurately predicted to within ten per cent.
Ding was initially unsure how meaningful his results were and was curious to hear Villaescusa-Navarro’s perspective. He was more than skeptical. “My first thought was, This is completely crazy, I don’t believe it, this is the work of an undergraduate, there must be a mistake,” Villaescusa-Navarro said. “I asked him to run the program in a few other ways to see if he would still come up with similar results.” The results held.
Villaescusa-Navarro began to do his own calculations. His doubt focussed foremost on the way that the machine learning itself worked. “One thing about neural networks is that they are amazing at finding correlations, but they also can pick up on numerical artifacts,” he said. Was a parameter wrong? Was there a bug in the code? Villaescusa-Navarro wrote his own program, to ask the same sort of question that he had assigned to Ding: What could information about one galaxy say about the universe in which it resided? Even when asked by a different program, written from scratch, the answer was still coming out the same. This suggested that the result was catching something real.
“But we couldn’t just publish that,” Villaescusa-Navarro said. “We needed to try and understand why this might be working.” It was working for small galaxies, and for large galaxies, and for galaxies with very different features; only for a small handful of eccentric galaxies did the work not hold. Why?
The recipe for making a universe is to start with a lot of hydrogen, a little helium, some dark matter, and some dark energy. Dark matter has mass, like the matter we’re familiar with, but it doesn’t reflect or emit light, so we can’t see it. We also can’t see dark energy, but we can think of it as working in the opposite direction of gravity. The universe’s matter, via gravity, pushes it to contract; the universe’s dark energy pushes it to expand.
Omega matter is a cosmological parameter that describes how much dark matter is in the universe. Along with other parameters, it controls how much the universe is expanding. The higher its value, the slower the universe would grow. One of the research group’s hypotheses to explain their results is, roughly, that the amount of dark matter in a universe has a very strong effect on a galaxy’s properties—a stronger effect than other characteristics. For this reason, even one galaxy could have something to say about the Omega matter of its parent universe, since Omega matter is correlated to what can be pictured as the density of matter that makes a galaxy clump together.
In December, Genel, an expert on galaxy formation, presented the preliminary results of the paper to the galaxy-formation group he belongs to at the Center for Computational Astrophysics, in New York. “This was really one of the most fun things that happened to me,” he said. He told me that any galaxy-formation expert could have no other first reaction than to think, This is impossible. A galaxy is, on the scale of a universe, about as substantial as a grain of sand is, relative to the size of the Earth. To think that all by itself it can say something so substantial is, to the majority of the astrophysics community, extremely surprising, in a way analogous to the discovery that each of our cells—from a fingernail cell to a liver cell—contains coding describing our entire body. (Though maybe to the poetic way of thinking—to see the world in a grain of sand—the surprise is that this is surprising.)
Rachel Somerville, an astrophysicist who was at the talk, recalled the initial reaction as “skepticism, but respectful skepticism, since we knew these were serious researchers.” She remembers being surprised that the approach had even been tried, since it seemed so tremendously unlikely that it would work. Since that time, the researchers have shared their coding and results with experts in the field; the results are taken to be credible and compelling, though the hesitations that the authors themselves have about the results remain.
The results are not “robust”—for now, the computer can make valid predictions only on the type of universe that it has been trained on. Even within CAMELS, there are two varieties of simulations, and, if the machine is trained on one variety, it cannot be used to make predictions for galaxies in the other variety. That also means that the results cannot be used to make predictions about the universe we live in—at least not yet.
Villaescusa-Navarro told me, “It is a very beautiful result—I know I shouldn’t say that about my own work.” But what is beauty to an astrophysicist? “It’s about an unexpected connection between two things that seemed not to be related. In this case, cosmology and galaxy formation. It’s about something hidden being revealed.”