Learning to Compress Plasma Turbulence

Gianluca Galletti, Gerald Gutenbrunner, Fabian Paischer, Sandeep Suresh Cranganore, William Hornsby, Naomi Carey, Lorenzo Zanisi, Stanislas Pamela, Johannes Brandstetter

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: turbulence, plasma, neural compression, autoencoders, neural fields

TL;DR: Neural compression methods enable extreme compression of plasma turbulence simulation data while maintaining low reconstruction error and preserving key physical characteristics.

Abstract:

Turbulence is a nonlinear phenomenon exhibiting chaotic, multiscale behavior. It is simulated with high-fidelity numerical solvers operating on fine grids, making the process both computationally demanding and storage intensive. A prime example is gyrokinetics, which simulates turbulence in a magnetized plasma.
A single run can take weeks and produce up to tens of terabytes of data, making storage unfeasible even with standard compression algorithms. Raw data is typically discarded, which is wasteful and prevents routine post-hoc analysis of the simulations.
To this end, we investigate neural compression methods capable of extreme compression ratios (up to 40,000x) while preserving reconstruction quality and physical fidelity.
Our study focuses on autoencoders and neural implicit fields, specifically trained with novel physics-informed losses.
This direction could enable practitioners to store high-fidelity turbulence simulations for downstream scientific analysis at a fraction of the volume.

Submission Number: 183