Can AI help us understand dark energy better?

Explaining the origins and evolution of the Universe has taxed the brightest brains the world has produced, like Cambridge’s Stephen Hawking. But what we’re rapidly realising - and Stephen Hawking himself alluded to  over the course of his career and his writings - is that the human brain may just not be capable of comprehending the multidimensional relationships that drive the cosmos. So scientists are increasingly turning to computers and machine learning to look for the patterns that might steer us along the right path. Roberto Trotta is one of them. He’s at the International School for Advanced Studies (SISSA). They occupy what was once a hospital high in the hills overlooking Trieste, so anyone who works there gets the brain-stimulating bonus of an incredible vista…

Chris - You are in theory a cosmologist, but you probably should have a degree in real estate because if you got this office with a view like that, you should be selling apartments.

Roberto - Well, in fact, they sold this office to me and in fact I was a very happy buyer every time I come in.

Chris - For listeners at home, we are sitting in Roberto's office and I'm staring over a balcony at the most beautiful seascape bay dotted with boats. What are we, a few miles up the hills here from the sea?

Roberto - Yeah, we're just about 300 meters high up on the beginning of the Karst plateau overlooking Trieste and we can see, on one side, the eastern peninsula. On the other side, on a good day, you can almost imagine seeing Venice shimmering in the distance.

Chris - Was this always an institute where big brained physicists probe the origins of the universe, this building?

Roberto - Well it actually used to be a hospital for a patient with tuberculosis. So they needed the air, the fresh air of the Karst plateau, and they needed the terrace, which is a relic of that epoch. So our offices enjoy the view and the fresh air on the terrace as well.

Chris - You're devoted to trying to understand the evolution of the universe. Can you try to capture for us about 13.8 billion years, how it all fits together and where the interesting bits are?

Roberto - So the universe starts in a big bang, a big hot energetic state where the universe starts expanding at that accelerated rate. And then all of this disappears, and light, matter, neutrinos, dark matter, all the stuff the university has made of, emerges.

Chris - What would be the timeline on that?

Roberto - Probably something around 10 to the minus 32 seconds, something like that. Very, very short amount of time. A fraction of a fraction of a second. After that, it's another three minutes before all the atoms actually come into existence in a way that we recognized today, mostly hydrogen and helium. And after that it's 380,000 years of opaqueness. The universe is filled with a plasma, which means it's a very hot radiation and hydrogen bath that is so dense that light itself cannot propagate. So it's literally a moment in the life of the universe where it's like looking through a fog. You can't see anything. But then the magic happens. The universe cools off sufficiently so that electrons get captured by protons. The universe becomes neutral. Now light can propagate and a beautiful map of the early universe emerges and that map we can pick up with our telescopes. So we can look all the way back to this moment in time, 380,000 years after the Big Bang.

Chris - And that was when the first stars that we can see were shining?

Roberto - Not quite. At that moment in time, stars haven't had time to form yet. We need to wait another 500 million years before the first stars and galaxies begin to form. And then gravity starts doing its job. Galaxies form and so on and so forth. And things continue pretty smoothly for about 6 billion years until something else happens.

Chris - How do you know it's that time point? How do you know that things were pretty business as usual until 6 billion years?

Roberto - The most important marker of that point in time is the fact that we can look back at the expansion of the universe. And we see the universe expanding but decelerating, slowing down as it expands. This is what we expect from gravity because gravity tends to bring things together. But 6 billion years ago, something strange happens and the expansion of the universe starts picking up speed again, powered by something mysterious that we're trying to find out.

Chris - And that's dark energy, that's pushing things apart.

Roberto - Dark energy, we don't know exactly what it is, but it's one of the biggest mysteries in physics today. We know or we think we know that 70% of the universe is made of dark energy, yet we don't know what it is.

Chris - Do we have a clue as to why it took 6 billion years before dark energy begins to dominate and starts making things grow faster? Was there just a sort of balancing act going on with gravity, the matter we could identify, and some dark energy, but then it reaches a sort of tipping point where there's enough there to enter almost like a positive feedback loop where the faster it goes, the faster it grows.

Roberto - That's right. The fundamental property of dark energy, if it is indeed the property of empty space itself, is that as the universe expands, we create more empty space and form more dark energy. By contrast, matter of any kind, whether normal matter or dark matter, as the universe expands, gets diluted, there is the same amount of matter, but more volume, more space, therefore the relative importance goes down. So there is a tipping point like you rightly said, where dark energy takes over because there is more space, there is more dark energy. But the deeper question is - why at that point? Why not at an earlier time or a later time in the history universe? That's the deeper question, which we really don't know how to answer yet.

Chris - Are we not in a position though where we are into a realm of science now, where what we're trying to consider, the amount of information we're trying to relate, is so vast that it is beyond the comprehension even of big brained cosmologists?

Roberto - Cosmology has been undergoing a revolution. In the sixties, cosmology was not considered a real science because it was a playground for theorists because there were hardly any observations, constraining any crazy idea. Nowadays we're almost at the opposite end. We have so many observations, so much data, understanding is the bottleneck now and how to go from data to understanding and from understanding to theoretical models and the four deeper insights into the nature of reality and the nature of the universe. That is what we are really banging our head against. The usual tools or the sort of classical tools that we've been using for data analysis and statistics, are not up to the task. And so we are now turning towards artificial intelligence and machine learning and modern tools that will hopefully be able to extract meaning, extract patterns, extract structure from a large amount of data that no astronomer or no cosmologist, no human being can ever hope to be able to analyze, let alone look at. So it's really a question of having the machines understand what's relevant for us and sort of summarize it in a way that we can then interpret in a physically hopefully meaningful way.

Chris - Everyone I talk to about AI and machine learning says to me, 'this is great. It can spot patterns between one piece of information and another'. But when you ask it, 'how did you win that game of Go'? The machine can't tell you. It's not explainable. So is this not a difficulty? Because you'll get answers out, but you won't necessarily get the why?

Roberto - Very much so, and that's one that we as physicists really are struggling with. For us we really want to understand what those structures, what those patterns mean and how the algorithms actually come to the conclusions that they come to. We've been really, really successful as a community in coming up with new ways of doing machine learning and AI that are slightly different from the Googles of this world, who have slightly different problems than us. They want to spot patterns in a mass of data for which there is no theory. We don't know why and how human beings behave, we just want to see what those patterns are. In our case, when we see galaxy distributions in the sky or whatnot, we have a theoretical understanding up to a point where those patterns come from and we can give some of this knowledge to the machine so that the machine can add on top things that we don't know about. So it's really a question of finding a sweet spot where we provide the machine with our human understanding and hopefully the machine can provide something else that we are missing at the moment. And so it's really finding that new equilibrium between machine and human intelligence that might solve the problems. If machine learning will discover a new property of dark energy or new patterns that we cannot explain where they come from, will we be able to trust it? Will we be able to say this is the true theory of the universe, if no human can ever explain where it comes from? That's very unsatisfactory. So we're working on making it more explainable, more understandable.