Artificial intelligence technology has ushered in a new era of high-resolution simulation of the universe

By using neural networks, researchers can now simulate the universe for a small amount of time, which greatly promotes the development of physics research. A universe has undergone billions of years of evolution, but researchers have developed a method to create a complex simulated universe in less than a day. This technology, recently published in the Proceedings of the National Academy of Sciences, combines machine learning, high-performance computing, and astrophysics to help usher in a new era of high-resolution cosmological simulations.

Cosmological simulations are an important part of uncovering many mysteries of the universe, including the mysteries of dark matter and dark energy. But until now, researchers still face the challenge of fish and bear paws, which is how to achieve simulations that can focus on a small area at high resolution or encompass a large universe at low resolution.

Tiziana Di Matteo and Rupert Croft, physics professors at Carnegie Mellon University, Yin Li, a researcher at the Flatiron Institute, Yueying Ni, a doctoral student at Carnegie Mellon University, Simeon Bird, physics and astronomy professor at the University of California, Riverside, and Yu Feng from the University of California, Berkeley overcame this problem by teaching a neural network-based machine learning algorithm to upgrade simulations from low resolution to super-resolution.

Tiziana Di Matteo and Rupert Croft, physics professors at Carnegie Mellon University, Yin Li, a researcher at the Flatiron Institute, Yueying Ni, a doctoral student at Carnegie Mellon University, Simeon Bird, physics and astronomy professor at the University of California, Riverside, and Yu Feng from the University of California, Berkeley overcame this problem by teaching a neural network-based machine learning algorithm to upgrade simulations from low resolution to super-resolution.

The trained code can use a full-size low resolution model and generate super-resolution simulations containing up to 512 times the size of particles. For an area of approximately 500 million light-years in the universe, containing 134 million particles, existing methods will take 560 hours to stir up a high-resolution model using a processing core, while using the new method, researchers will only need 36 minutes.

When more particles are added to the simulation, the results are even more dramatic. For a universe 1000 times larger with 134 billion particles, researchers' new method takes 16 hours on a single graphics processing unit. Using traditional methods to complete simulations of this scale and resolution will require a dedicated supercomputer.

Scientists use cosmological simulations to predict the appearance of the universe in various situations. For example, if the dark energy in the universe changes over time and is then observed through a telescope to confirm whether the simulated predictions are realistic.

"The universe is a big dataset - artificial intelligence is the key to understanding the universe and revealing new physics," said Scott Dodelson, professor and director of the Department of Physics at Carnegie Mellon University and director of the National Science Foundation's Planning and Research Institute. "This study demonstrates how the National Science Foundation's Artificial Intelligence Planning and Research Institute will promote the development of physics through artificial intelligence, machine learning, statistics, and data science."

"It is evident that artificial intelligence is having a huge impact on many scientific fields, including physics and astronomy," said James Shank, project director of the NSF Physics Department. "Our AI Planning and Research Institute project is working hard to accelerate AI discovery. This new achievement is a great example of how AI can change cosmology."

Ni and Li used these fields to create a set of code that uses neural networks to predict how gravity moves dark matter over a period of time. These networks receive training data, run calculations, and compare the results with expected results. With further training, the network adapts and becomes more accurate.

The specific method used by the researchers is called generative adversarial networks, where two neural networks compete against each other. A network simulates the universe at low resolution and uses them to generate high-resolution models. Another network attempts to distinguish these simulations from those made using traditional methods. As time passed, both neural networks became better and better, until in the end, the simulation generator emerged victorious and created the ability to simulate quickly.

Although trained using only a small area of space, neural networks accurately replicate large-scale structures that only appear in large-scale simulations.

However, these simulations did not capture everything. Because they focus on dark matter and gravity, smaller scale phenomena such as star formation, supernovae, and black holes are excluded. The researchers plan to expand their methods to include forces responsible for such phenomena and load their neural networks in conventional simulations to improve accuracy.

Prev: X-NAND technology is expected to bring SLC level performance to QLC flash memory