Description
We present jf1uids (https://github.com/leo1200/jf1uids), a differentiable magnetohydrodynamics (MHD) simulator written in JAX, scaling to multiple GPUs. Our open source code features a modern inherently divergence free approach to MHD, (near-)energy-conserving self-gravity stable at discontinuities, a conservative geometric formulation for radially symmetric simulations and multiple physics modules in development, starting with cosmic rays and stellar winds. It's differentiability and scalability make jf1uids a unique tool for bringing together classical astrophysical simulations and machine learning. Being written in high-level Python / JAX, jf1uids allows for rapid development and contributions from a broad community.
Primary author
Leonard Storcks
(Heidelberg University)
Co-author
Dr
Tobias Buck
(Heidelberg University)
