Aug 30 2019
As part of a new study, scientists have successfully used artificial intelligence to develop a model of the Universe.
A comparison of the accuracy of two models of the Universe. The new deep learning model (left), dubbed D3M, is much more accurate than an existing analytic method (right) called 2LPT. The colors represent the error in displacement at each point relative to the numerical simulation, which is accurate but much slower than the deep learning model. (Image credit: S. He et al./PNAS2019)
In general, scientists strive to gain insights into the Universe by making model predictions to match observations. In the past, they could model simple or highly simplified physical systems, humorously termed the “spherical cows,” with pencils and paper.
Then, the advent of computers allowed them to model complicated phenomena using numerical simulations. For instance, scientists have programmed supercomputers to mimic the motion of billions of particles through billions of years of cosmic time, a process called the N-body simulations, to explore how the Universe evolved to how it looks at present.
“Now with machine learning, we have developed the first neural network model of the Universe, and demonstrated there’s a third route to making predictions, one that combines the merits of both analytic calculation and numerical simulation,” stated Yin Li, a Postdoctoral Researcher at the Kavli Institute for the Physics and Mathematics of the Universe, University of Tokyo, and jointly the University of California, Berkeley.
At the start, things in the Universe were highly uniform. With time, due to gravity, the denser parts became denser and sparser parts turned sparser, ultimately paving the way to a foam-like structure called the “cosmic web.” Researchers have used various techniques to analyze this structure formation process, such as numerical simulations and analytic calculations.
Although analytic techniques are fast, they fail to generate precise results for large density fluctuations. By contrast, numerical (N-body) techniques mimic the structure formation precisely; however, it is very expensive to track gazillions of particles, even on supercomputers. Hence, researchers usually face the efficiency versus accuracy trade-off while trying to model the Universe.
However, the rapid growth of observational data in terms of quantity and quality necessitates techniques that excel in both efficiency and accuracy.
In order to overcome this difficulty, a team of scientists from the United States, Canada, and Japan, including Li, focused on machine learning, an advanced technique for the detection of patterns and for making predictions. Similar to machine learning that has the ability to transform the portrait of a young man into his older self, Li and his team questioned whether it could also estimate how universes evolve based on their early snapshots.
They developed a deep learning model with the potential to simulate the structure formation process by training a convolutional neural network with simulation data of trillions of cubic light-years in volume. The new model is not just several times more precise than the analytic techniques, it is also considerably more efficient when compared to the numerical simulations used for its training.
It has the strengths of both previous [analytic calculation and numerical simulation] methods.
Yin Li, Postdoctoral Researcher, Kavli Institute for the Physics and Mathematics of the Universe, University of Tokyo
According to Li, the power of AI emulation will scale up in the years to come. N-body simulations are already strongly optimized, and as a maiden effort, his group’s AI model still has a huge scope for improvement.
In addition, more complex phenomena are more expensive to simulate, but not possible so on emulation. Li and his team anticipate a higher performance gain from their AI emulator upon moving on to incorporating other effects, like hydrodynamics, into the simulations.
It won’t be long before we can uncover the initial conditions of and the physics encoded in our Universe along this path.
Yin Li, Postdoctoral Researcher, Kavli Institute for the Physics and Mathematics of the Universe, University of Tokyo
This research work was published on June 24th, 2019, in the Proceedings of the National Academy of Sciences of the United States of America.
Source: https://www.ipmu.jp/en