INTERESTS研究紹介

Surrogate modeling for Supernova feedback

Supernovae (SNe) are the explosions at the end of the life of massive stars. They are one of the most important sources of energy and momentum, which drive the evolution of galaxies. However, simulating the feedback from SNe is computationally expensive because it requires resolving the multiscale physical processes with tiny timesteps. In response, we have developed a surrogate model using machine learning to duplicate supernova feedback quickly. When a supernova explodes in a galaxy simulation, the model duplicates the SN feedback 100,000 years ahead on-the-fly. This approach accelerates the simulation by a factor of 100 compared to the direct simulation. (Hirashima et al. 2023a,b)

Read more

See more

PHEASANT: Exascale Star-by-star Galaxy Simulations

To investigate the multi-phase structure of interstellar medium and circulation of matter in galaxies, we are developing a new galaxy formation simulation code that can resolve individual stars. We are applying deep learning and image processing techniques to accelerate the simulation. The PHEASANT collaboration has extended this approach to architecture-agnostic exascale simulations of a Milky Way-mass galaxy, reaching 5.4 trillion particles across Intel, AMD, and NVIDIA GPU systems. The associated SC26 paper was selected as an ACM Gordon Bell Prize Finalist. (Hirashima et al. 2025, 2026)

See PHEASANT

Image credit: NASA/JPL-Caltech

Foundation model for astro/fluid simulations

Cosmological/galaxy simulations have many physical processes that need to be resolved, such as supernova and AGN feedback. As a part of PolymathicAI initiative, I am developing a foundation model that can be used for various hydrodynamical simulations in astrophysics. The model is based on generative deep learning architectures that can be trained on various physical simulations.

(Polymathic AI initiative)

See more