Title: Extracting Collision Operators from Plasma Phase Space Data via Differentiable Simulators

Abstract:

This study introduces a novel methodology for deriving collision operators from phase space data generated by plasma dynamics. The proposed framework integrates a gradient-based optimization technique with a differentiable kinetic simulator, specifically employing a differentiable Fokker-Planck solver as its central element. This combination enables the learning of collisional operators that most accurately characterize the observed phase space evolution.

We validated our approach using data obtained from two-dimensional Particle-in-Cell (PIC) simulations of spatially uniform thermal plasmas. Through this process, we successfully learned the collision operator that accounts for the self-consistent electromagnetic interactions among finite-size charged particles across a broad spectrum of simulation parameters. Our findings indicate that these learned operators outperform alternative estimates derived from particle tracks. Furthermore, this method eliminates the need for prior assumptions regarding the relevant time scales of the underlying processes and substantially lowers memory consumption.

In the non-relativistic regime, the operators retrieved by our method align closely with theoretical predictions established for electrostatic scenarios. These results highlight the potential of differentiable simulators as a computationally efficient and powerful tool for inferring new operators applicable to a diverse array of challenges, including electromagnetically dominated collisional dynamics and stochastic wave-particle interactions.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC