CDS&E: Data-driven fast methods for high-energy plasma astrophysics

University Of California-Santa Cruz, Santa Cruz CA

Investigators

Abstract

Significantly energetic astrophysical flows exhibiting plasma velocities near the speed of light are referred to as relativistic plasmas. These high-energy relativistic flows manifest in a diverse array of astrophysical phenomena involving compact objects. Noteworthy examples encompass core-collapse supernovae, jets and accretion flows surrounding massive compact objects like black holes and neutron stars, pulsar wind nebulae, and gamma-ray bursts. Moreover, astronomical observations consistently indicate the presence of dynamically significant magnetic fields within these highly compressible relativistic flows. Computer simulations serve as indispensable tools for researchers studying these plasmas, facilitating the comprehension of various physical processes associated with immensely energetic relativistic astrophysical flows. Scientists at the University of California, Santa Cruz aim to advance the precision of computer simulations concerning relativistic flows, as current computer algorithms encounter challenges in delivering high-fidelity, dependable numerical solutions. The team will develop a novel data-driven machine-learning strategy that enhances the computational performance and accuracy of numerical solutions for relativistic plasma flows. The expected outcomes of this project will be disseminated to the broader computational astrophysics community through publications in scientific journals and the release of open-source code for improved computer simulations. As part of this project, the PI will also advise and mentor undergraduate students from underrepresented groups via the Cal-Bridge and Lamat REU programs. The project's primary objective is to tackle unresolved challenges in simulating relativistic flows. Currently, simulating relativistic flows using modern shock-capturing schemes necessitates an impractically high grid resolution to achieve grid convergence. To address this issue, the team proposes the development of new data-driven, fast, a-priori shock-capturing methods for relativistic hydrodynamics and magnetohydrodynamics. The proposed approach aims to eliminate the need for computationally expensive conventional "limited reconstruction" of fluid data, which is a nonlinear numerical process required for numerical stability in standard modern shock-capturing methods. To overcome this limitation, the team will create high-order "unlimited" reconstruction algorithms using Gaussian Process (GP) reconstruction. By combining the GP method with a physics-informed artificial neural network, they will introduce a novel data-learned shock-capturing paradigm named the a-priori annMOOD (Artificial Neural Network Multidimensional Optimal Order Detection) method, which will replace the existing a-posteriori procedural shock-capturing MOOD method. The anticipated outcome of this project is a performance-enhanced relativistic (magneto)hydrodynamics (RMHD) solver optimized for massively parallel computing. By leveraging the power of data-driven techniques and high-order reconstruction algorithms, this project aims to significantly improve the efficiency and accuracy of simulating relativistic flows, thereby advancing understanding of these complex phenomena. The resulting solver will be capable of delivering reliable results while reducing the computational burden associated with achieving grid convergence. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.