The researchers created a first-ever artificial intelligence (AI) simulation of the universe that seems to work like a real thing.
Mysterious 3D AI Simulation of Universe
In the journal Proceedings of the National Academy of Sciences, the researchers have reported a new simulation that was intended to create a virtual version of the cosmos and simulate different conditions for the beginning of the universe. However, it turned out to work so well that the scientists are curious to study their own simulation and understand how is it working so well.
One of the authors, Shirley Ho, stated that the AI simulation is working like an image-recognition software that is taught to recognize pictures of cats and dogs. But suddenly, it’s able to recognize elephants. No one knows how it happened and it has become a great mystery to be solved.
How Scientists Simulate the Universe?
Studying the formation of Universe is always a challenging task for researchers because of the enormous age and scale of the universe. For this purpose, astrophysicists use computer modeling which requires a lot of computing time and power.
In the conventional method, they might need to run thousands of simulations, tweaking no. of parameters to determine which is the most likely real-world scenario.
When Shirley Ho and her co-researchers decided to make a simulation of the universe, they created a deep neural network to speed up the process. For this purpose, they designed and dubbed the neural network as Deep Density Displacement Model or to recognize common features in data.
The main purpose of this study was to learn over time how to manipulate that data. In the case of , the researchers used 8,000 simulations as input from a high-accuracy traditional computer model of the universe. Once they learned how those simulations worked, they put in a brand-new, simulation of a virtual, cube-shaped universe that was 600 million light-years across. It meant that the simulation was never seen before by the scientists.
The neural network was able to run simulations in the new universe just like it had a dataset of 8,000 simulations dataset that it used for training. The simulations mainly focused on the role of gravity in the formation of the universe.
What was surprising for the researchers was that when they used to vary brand-new parameters, like the amount of dark matter in the virtual universe, was still able to handle the simulations. Despite the fact that the network was never being trained on how to handle dark matter variations, it was working like a real-time scenario.
This astonished the researchers very much. According to them, feature is a mystery that makes the simulation intriguing for computational science as well as cosmology.
The new neural network is so fast that it could complete simulations in 30 milliseconds, compared to several minutes for the fastest non-AI simulation method. Moreover, it had an error rate of 2.8%, compared with 9.3% error rate for the existing fastest model.
The researchers are now planning to vary other parameters in the new neural network to examine how factors like hydrodynamics or the movement of fluids and gases, may tweak the formation of the universe.
This new simulation can be an exciting field for machine learners to study why this model extrapolates real-life scenarios so well and how it recognizes elephants instead of just recognizing cats and dogs. It’s just like a two-way street between science and deep learning. This model might also act as a time-saver for researchers who are interested in studying universal origins.
Study Reference: https://www.pnas.org/content/early/2019/06/21/1821458116