Afterlive: A performant code for Vlasov-Hybrid simulations

TL;DR

Afterlive introduces a highly parallelized, reconstruction-based Vlasov-Hybrid simulation code that effectively captures phase space dynamics with a larger dynamic range, improving accuracy over traditional particle methods.

Contribution

The paper presents a novel, parallelized implementation of the Vlasov-Hybrid method that reconstructs the distribution function at each step, enabling larger dynamic ranges and improved phase space representation.

Findings

Efficient parallelization using MPI and HDF5.

Enhanced dynamic range in phase space density.

Accurate self-consistent electrostatic field calculation.

Abstract

A parallelized implementation of the Vlasov-Hybrid method [Nunn, 1993] is presented. This method is a hybrid between a gridded Eulerian description and Lagrangian meta-particles. Unlike the Particle-in-Cell method [Dawson, 1983] which simply adds up the contribution of meta-particles, this method does a reconstruction of the distribution function $f$ in every time step for each species. This interpolation method combines meta-particles with different weights in such a way that particles with large weight do not drown out particles that represent small contributions to the phase space density. These core properties allow the use of a much larger range of macro factors and can thus represent a much larger dynamic range in phase space density. The reconstructed phase space density $f$ is used to calculate momenta of the distribution function such as the charge density $\rho$. The charge density $\rho$ is also used as input into a spectral solver that calculates the self-consistent electrostatic field which is used to update the particles for the next time-step. Afterlive (A Fourier-based Tool in the Electrostatic limit for the Rapid Low-noise Integration of the Vlasov Equation) is fully parallelized using MPI and writes output using parallel HDF5. The input to the simulation is read from a JSON description that sets the initial particle distributions as well as domain size and discretization constraints. The implementation presented here is intentionally limited to one spatial dimension and resolves one or three dimensions in velocity space. Additional spatial dimensions can be added in a straight forward way, but make runs computationally even more costly.