Kinetic Simulations of Space Plasmas Using Recurrent Neural Networks
||Kinetic Simulations of Space Plasmas Using Recurrent Neural Networks|
||Shahab Fatemi <email@example.com>|
||2023-10-01 – 2024-04-01|
We have developed the Amitis code, a three-dimensional, multi-species, time-dependent hybrid-kinetic model of plasma (kinetic ions and fluid electrons) that runs in parallel on multiple GPU nodes. The model is well-established within the space physics community and since its development in 2017, the model has directly contributed to 15 peer-reviewed scientific publications . Thanks to the previous support by NSC/Berzelius (project SNIC Berzelius-2022-177), we have recently upgraded Amitis by adding a new feature into it that uses Recurrent Neural Networks (RNN) to solve the equation of motion for plasma. This new feature is tested and operational on our model based on the problem of our interest. We have developed this method as part of the particle solver package of our massively parallel kinetic model of plasma, Amitis. In general, Amitis performs well on A100 GPUs, and the model utilization is on average over 85% using a single node, and over 70% using 8 nodes on Berzelius.
The goal of this computation project is to continue using Berzelius, and together with our advanced kinetic model of plasma, we are going to provide a detailed understanding on the interaction between plasma and different planetary bodies inside and outside our solar system. Our primary focus is on the global and local structures of plasma interactions with outer planetary moons (e.g., Ganymede, Europa, and Callisto) and the stellar wind interaction with terrestrial bodies (e.g., Mercury, Earth, and Exoplanets). We have recently received two project grants to hire 2 PhD students at Umeå University. These projects are on average 75% simulation based using the Amitis model where GPUs are essential for running the model. We have been using Bezelius before and continuing your support is critical to move our research forward.