Nonlinear kinetic closures for linear instabilities

Landau fluid closures are linear in the evolved moments. We examine the consequences of relaxing linearity. As an example, we introduce the nonlinear exact kinetic response closure, exact on an eigenmode but not on superpositions. A counting argument shows that a closure on $N$ instantaneous moments and the field can represent at most $(N+1)\mathrm{/}2$ superposed eigenmodes. Guided by this, we train an equivariant neural-network and a learned Padé closure on analytic data. For bump-on-tail and slab ITG instabilities they reduce growth-rate errors by up to three orders of magnitude over Landau closures.