For the first time, astrophysicists used artificial intelligence techniques to create complex 3D simulations of the universe. The results are so fast, accurate and robust that even the developers are not sure how it all works.
"We can do these simulations in a few milliseconds, while other 'fast' simulations take a few minutes," says study co-author Shirley Ho, group leader at the Flatiron Institute's Center for Computational Astrophysics in New York City, and an extraordinary one Professor at Carnegie Mellon University. "Not only that, we are much more accurate."
The speed and accuracy of the project, abbreviated D 3 M, was not the biggest surprise for the researchers. The real shock was that D 3 M could accurately simulate what the universe would look like if certain parameters were adjusted – how much dark matter in the cosmos – although the model had never received training data, these parameters varied.
"It's like teaching image recognition software with lots of pictures of cats and dogs, but then you can spot elephants," explains Ho. "No one knows how to do that, and it's a big puzzle to be solved . "
Ho and her colleagues present D 3 M June 24 in Proceedings of the National Academy of Sciences . The study was led by Siyu He, a research analyst at the Flatiron Institute.
Ho and He collaborated with Yin Li of the Berkeley Center for Cosmological Physics at the University of California at Berkeley and the Kavli Institute of Physics and Mathematics of the Universe near Tokyo; Yu Feng from the Berkeley Center for Cosmological Physics; Wei Chen from the Flatiron Institute; Siamak Ravanbakhsh from the University of British Columbia, Vancouver; and Barnabás Póczos of Carnegie Mellon University.
Computer simulations such as those of D 3 M have become indispensable for theoretical astrophysics. Scientists want to know how the cosmos could evolve under different scenarios, for example, when the dark energy that pulls the universe apart varies over time. Such studies require the implementation of thousands of simulations, making a lightning fast and highly accurate computer model one of the main goals of modern astrophysics.
D 3 M models of how gravity shapes the universe. Researchers chose to concentrate only on gravity, as it is by far the most important force for the cosmos' evolution.
The most accurate universe simulations calculate how gravity shifts every billion individual particles across the earth throughout the universe's age. This accuracy is time-consuming and requires around 300 computation hours for a simulation. Faster methods can finish the same simulations in about two minutes, but the required joins result in lower accuracy.
Ho, He and their colleagues refined the deep neural network that powers D 3 M by feeding it with 8,000 different simulations of one of the most accurate models on the market. Neural networks capture training data and perform calculations based on the information. The researchers then compare the resulting result with the expected result. As training progresses, neural networks adapt over time for faster, more accurate results.
After the training of D 3 M, the researchers performed simulations of a box-shaped universe that was 600 million light-years long, and compared the results to those of the slow and fast models. While the slow but accurate approach took hundreds of hours of computation per simulation and the existing fast method took several minutes, D 3 M could perform a simulation in just 30 milliseconds.
D 3 M also provided accurate results. Compared to the highly accurate model, D 3 M had a relative error of 2.8 percent. Using the same comparison, the existing fast model had a relative error of 9.3 percent.
D 3 M's remarkable ability to handle parameter variations not found in his training data makes it a particularly useful and flexible tool. Ho says. Ho's team not only wants to model other forces such as hydrodynamics, but also to learn more about how the model works under the hood. This may have a positive effect on the promotion of artificial intelligence and machine learning, says Ho.
"For a machine learner, it can be an interesting playground to see why this model extrapolates so well and why it's not extrapolated to elephants Cats and dogs recognize, "she says. "It's a two-way street between science and deep learning."
CosmoGAN: Training a Neural Network to Study Dark Matter
Siyu He et al., Learning to Predict Cosmological Structure Formation, Proceedings of the National Academy of Sciences (2019). DOI: 10.1073 / pnas.1821458116
The first AI universe simulation is fast and accurate – and the developers do not know how it works (2019, June 26)
retrieved on June 26, 2019
This document is subject to copyright. Apart from any fair dealings for the purposes of private study or research, no
Part may be reproduced without written permission. The content is provided for informational purposes only.