AI vs. supercomputers, Round 1: galaxy simulation goes to AI

In the first study of its kind, researchers at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS) in Japan, together with colleagues from the Max Planck Institute for Astrophysics (MPA) and the Flatiron Institute, used machine learning — a form of artificial intelligence — to significantly speed up the processing time to simulate the evolution of galaxies coupled with supernova explosions. This approach could help us to understand the origins of our own galaxy and, in particular, the elements essential for life in the Milky Way.

Understanding how galaxies form is a key issue for astrophysicists. While we know that powerful events such as supernovae can influence the evolution of galaxies, it is not possible to observe this process directly by looking at the night sky. Instead, scientists rely on numerical simulations based on large amounts of data collected from telescopes and other devices that measure aspects of interstellar space. These simulations must account for gravity and hydrodynamics, as well as other complex aspects of astrophysical thermo-chemistry.

Additionally, they must have a high temporal resolution, meaning the time interval between each 3D snapshot of the evolving galaxy must be short enough to capture critical events. For instance, capturing the initial phase of supernova shell expansion requires a timescale of just a few hundred years, which is 1,000 times shorter than that achievable in typical interstellar space simulations. In fact, it takes a typical supercomputer 1-2 years to carry out a simulation of a relatively small galaxy at the proper temporal resolution.

Overcoming this time step bottleneck was the main goal of the new study. By incorporating AI into their data-driven model, the research group was able to match the output of a previously modelled dwarf galaxy, but much more quickly. 'When we use our AI model, the simulation is about four times faster than a standard numerical simulation,' says Keiya Hirashima at RIKEN. 'This corresponds to a reduction of several months to half a year's worth of computation time. Critically, our AI-assisted simulation was able to reproduce the dynamics important for capturing galaxy evolution and matter cycles, including star formation and galaxy outflows.”

As with most machine learning models, the researchers' new model was trained using one set of data and was then able to predict outcomes based on a new set of data. In this case, the model incorporated a programmed neural network and was trained using 300 simulations of an isolated supernova in a molecular cloud with a mass of one million suns. After training, the model could predict the density, temperature and three-dimensional velocities of the gas 100,000 years after a supernova explosion. Compared with direct numerical simulations, such as those performed by supercomputers, the new model produced similar structures and star formation histories, but took four times less time to compute. The lab is currently using the new framework to run a simulation of a Milky Way-sized galaxy.

'While tremendous progress has been made in using Machine Learning and Artificial Intelligence models to analyse a plethora of datasets across fields in physics, their use within the astrophysical community has been limited, mostly to studies of turbulence,' points out Ulrich Steinwandel at MPA. 'For the first time, we demonstrate that AI can be utilised to speed up a multi-scale physics problem with multiple feedback channels. Moreover, the methodology can capture the global star formation and outflow properties of simulated galaxies, as well as the detailed phase structure of these multiphase flows.'