Neural networks as a lossy compression and restart/recovery strategy for high-dimensional plasma simulations

High-resolution simulations, such as those of plasma dynamics in the Earth's magnetosphere, generate vast amounts of data that challenge storage and processing capabilities, especially in the Exascale computing era. Managing these data volumes requires innovative solutions that balance compression efficiency with the accuracy required for scientific analysis. In this work, we present a novel approach for lossy runtime data compression tailored for high-dimensional velocity distribution functions (VDFs) using the Vlasiator code. Our method integrates implicit neural representations, specifically Multi-Layer Perceptrons (MLPs), to train on and compress VDFs dynamically during the simulation runtime. We address the challenges of preserving physical correctness while achieving high compression ratios by employing techniques such as Fourier feature encoding and sinusoidal activation functions to mitigate spectral bias. We compare our method with industry-standard methods such as ZFP and demonstrate significant improvements in compression efficiency without compromising physical correctness.