A group of researchers from the Flatiron Institute’s Center for Computational Astrophysics has developed the world’s first deep learning-based 3D simulation of the universe. For the purpose of getting a better understanding of our universe, the researchers designed and developed a tool which can successfully simulate how the universe is and how it would change with respect to some physics parameters.
In their study, published in the journal PNAS (Proceedings of the National Academy of Sciences), researchers detail how they created a model of the universe using a deep neural network. Their model, called Density Displacement Model (D3M), was fed with more than 8000 simulations of the universe in order to learn how gravity shaped the universe as the main force. All of these 8000 simulations were created using highly-accurate universe simulator.
The training of the model took more than 300 hours, but for inference, the model takes only approximately 30 ms. This means researchers were able to generate 1000 simulations in less than 20 seconds.
Using the trained model, researchers ran simulations of a box-shaped universe, wide 600 million light years. They showed that the model outperforms existing models in the amount of processing time needed to generate a simulation, with a relative error of 2.8 percent which is also a large improvement over other methods.
More about the first neural network-based universe simulator can be read in the paper published in PNAS.
