Abstract: |
Recent galaxy formation simulations implemented with the N-body/SPH method, moving mesh approach, and adaptive mesh refinement. However, achieving sufficient parallelization efficiency presents challenges. Specifically, to represent a Milky Way-sized galaxy, even with the Zoom-in simulation, approximately 1 billion particles are required, and the mass resolution is limited to around 1,000 solar masses (Applebaum et al. 2021). To overcome these limitations, we are developing a new code that leverages the supercomputer "Fugaku" to simulate down to individual stars within the galaxy. Such codes employ hierarchical individual time-stepping methods, which can increase computational and communication costs by several hundred times. Thse are caused by short timescale phenomena, such as supernova explosions, leading to parallelization efficiency issues. As a solution to this problem, we are developing a surrogate model using machine learning to rapidly duplicate supernova feedback. This model learns from the results of supernova explosion simulations within giant molecular clouds and predicts the physical quantities a typical 100,000 years later. I report the fidelity of this technique in this presentation.
|