The Machine Learning Approach to Moment Closure Relations for Plasma: A Review

A new review catalogs a surge in machine-learning techniques aimed at solving a persistent bottleneck in plasma physics: the closure problem in fluid simulations. The analysis, posted on arXiv, surveys two main methodological families and outlines key hurdles that remain before these models can be reliably integrated into large-scale global simulations. The review, authored by Sam Burles and last revised on 15 June 2026, examines how machine learning is being deployed to develop closure models that can capture kinetic phenomena within plasma fluid frameworks [1]. Large-scale global simulations of plasma, essential for both space and laboratory physics, require a closure relation for high-order plasma moments, a long-standing computational challenge [2]. The paper organizes recent studies into two primary categories: neural-network surrogates and equation-discovery methods [1]. Neural-network approaches range from multilayer perceptrons to Fourier neural operators; the latter have recently demonstrated the ability to reproduce both linear and non-linear Landau damping online within a fluid solver [2]. Equation-discovery techniques, such as sparse regression, represent the second major methodological family under review [1]. The work further classifies studies by their testing environment, distinguishing between offline validation against reference data and online testing within a time-evolving solver [2]. This distinction is critical, as performance in a static dataset does not guarantee stability when a model is coupled to a live simulation. The review identifies several persistent obstacles. Off-diagonal pressure-tensor accuracy remains a significant issue, as does the ability of machine-learning closures to generalize beyond the distribution of data on which they were trained [1]. Stable integration into large-scale simulations is flagged as another central challenge that future research must address [2]. The paper appears on arXiv, an open-access repository that hosts preprints across physics, mathematics, and computer science and has grown to a submission rate of about 24,000 articles per month as of late 2024 [6]. The machine-learning architectures surveyed in the review share a lineage with broader advances in the field, including the transformer model introduced in the 2017 paper “Attention Is All You Need,” which has since become foundational to large language models and other AI systems [8]. By compiling and analyzing the rapid expansion of machine-learning closure methods, the review provides a structured snapshot of a subfield that sits at the intersection of computational plasma physics and modern data-driven modeling [1][2].