The universe within reach of Artificial Intelligence? Astrophysicists have succeeded in generating for the first time complex 3D simulations of the universe using AI techniques.
Fast and accurate universe simulations like never before, all thanks to Artificial Intelligence techniques used by astrophysicists at the Flatiron Institute’s Center for Computational Astrophysics in New York. But don't ask them too much either, because they seem themselves overwhelmed by their own babies' functioning!
On June 24 at the National Academy of Sciences, Siyu He, research analyst at the Flatiron Institute, and Shirley Ho, co-author of the study, group leader at the same Flatiron Institute’s Center for Computational Astrophysics and adjunct professor at Carnegie Mellon University, presented their new project called D3M for Deep Density Displacement Model. A collaborative effort with other scientists such as their colleague from the Flatiron Institute Wei Chen , Yu Feng and Yin Li from the Berkeley Center for Cosmological Physics at the University of California, Berkeley, and the Kavli Institute for the Physics and Mathematics of the Universe in Kashiwa, near Tokyo, for the second, Barnabás Póczos from Carnegie Mellon University in Pittsburgh or Siamak Ravanbakhsh from the University of British Columbia in Vancouver.
The AI can now simulate the universe...
Their new model is a real breakthrough in the world of astrophysics, which has always sought to combine speed and precision in the generation of simulations. For comparison purposes and to fully appreciate the performance achieved by the D3M, a simulation of the universe as accurate as possible may require some 300 hours of calculation compared to a few minutes for the least accurate.
The D3M simulates the universe in 30 milliseconds ! For what precision, you would ask me? For a relative error of 2.8% compared to the most accurate models (300 hours of calculation therefore). The fastest model available (a few minutes of calculation) had until then a relative error of 9.3% compared to the most accurate ones. "We can run these simulations in a few milliseconds, while other fast simulations take a few minutes, Ho confirms. Not only that, but we're also much more accurate."
8,000 simulations to train a neural network
How did they achieve such a feat? By feeding the beast first, or at least Artificial Intelligence. Siyue He, Shirley Ho and their team first fed the deep neural network with 8,000 different simulations from one of the most accurate models, reports the Simons Foundation in a article dated of the last June 26th (The First AI Simulation of the Universe Is Fast and Accurate - and Even Its Creators Don't Know How It Works). The D3M then took this data to perform its own calculations and push its drives further and further until the final simulation of a 600 million light-year universe in the form of a box (see diagram above). This simulation of the universe is based on gravity, the force that most influences the evolution of the cosmos.
But the D3M did not only perform simulations of the universe, combining precision and speed. Artificial Intelligence has thus succeeded in other simulations of the universe by taking into account the modification of certain parameters, such as the quantity of dark matter for example, without ever having had such data on the subject to train before. The deep neural network thus exceeded all expectations.
But that is paradoxically where the problem lies. No matter how fast, accurate and solid it may be, the D3M leaves its creators wanting more. They are not yet able to answer some questions about their machine's own operation!
Hence to think that Artificial Intelligence could in this case exceed the limits of the human being? It is only a step that some people, in the name of ethics, would be ready to take. "It's like training image recognition software with a lot of pictures of cats and dogs and then being able to recognize elephants, Ho explains. No one knows how he does it, it's a big mystery to solve." She concluded: "We can be an interesting playground for machine learners using our model to see why it extrapolates so well, going so far as to recognize elephants instead of just recognizing cats and dogs".