To attain a better understanding of the cosmos, researchers successfully developed the first deep learning-based 3D simulation of the universe.
“We can run these simulations in a few milliseconds, while other ‘fast’ simulations take a couple of minutes,” says study co-author Shirley Ho, a group leader at the Flatiron Institute’s Center for Computational Astrophysics in New York City and an adjunct professor at Carnegie Mellon University. “Not only that, but we’re much more accurate.”
Working in collaboration with researchers from UC Berkeley, the Kavli Institute for the Physics and Mathematics of the Universe in Japan, University of British Columbia, and Carnegie Mellon University, the team developed a model called Deep Density Displacement Model (D3M). The tool accurately simulated how the universe looked and how it would look if certain parameters were tweaked.
Using an NVIDIA V100 GPU, with the cuDNN-accelerated PyTorch deep learning framework, the team first trained a deep neural network by feeding it 8,000 different simulations.
By using the GPU for inference, the researchers reduced the time to generate 1,000 simulations from 115 seconds to 20.
Once trained, the team ran simulations of a box-shaped universe 600 million light-years across, comparing the results to two models, a slow-but-accurate approach, a fast simulation, and the proposed method. The proposed model shattered previous records, outperforming them by hundreds of hours, and minutes, respectively, completing a simulation in just 30 milliseconds.
In terms of accuracy, the model had a relative error or 2.8 percent, much better than the existing fast model that has an error of 9.3 percent.
The simulator will help astrophysicists study how the cosmos might evolve under various conditions.
“D3M learns to predict cosmic structure in the nonlinear regime more accurately than our benchmark model 2LPT,” the researchers stated in their paper.
“Our model generalizes well to test simulations with cosmological parameters significantly different from the training set. This suggests that our deep-learning model can potentially be deployed for a range of simulations beyond the parameter space covered by the training data.”
The researchers have released the source code and training data of their implementation on GitHub.