Predicting the Expansion of Supernova Shells for High-Resolution Galaxy Simulations Using Deep Learning

Abstract

Small integration timesteps for a small fraction of the particles become a bottleneck for future galaxy simulations with a higher resolution, especially for massively parallel computing. As we increase the resolution, we must resolve physics on a smaller timescale while the total integration time is fixed as the universe age. The small timesteps for a small fraction of the particles worsen the scalability. More specifically, the regions affected by supernovae (SN) have the smallest timestep in the whole galaxy. Using a Hamiltonian splitting method, we calculate the SN regions with small timesteps using a few thousand CPU cores but integrate the entire galaxy using a shared timestep. For this approach, we need to pick up particles in regions, which will be affected by SN (the target particles) by the next global step (the integration timestep for the entire galaxy) in advance. In this work, we developed the deep learning model to predict the region where the shell due to a supernova explosion expands during one global step. In addition, we identify the target particles using image processing of the density distribution predicted by our deep learning model. Our algorithm could identify the target particles better than the method based on the analytical solution. This particle selection method using deep learning and the Hamiltonian splitting method will improve the performance of galaxy simulations with extremely high resolution.