Behavioral Evidence for Neuromorphic Threshold Management of ELM-like Instability in BOUT++

Abstract

Edge Localized Modes (ELMs) are not merely a fusion nuisance. They are a test of whether an intelligent controller can survive
contact with a violent nonlinear system before the system tears through a stability boundary. This preprint reports a behavior level evaluation of a neuromorphic model in the BOUT++ elm_pb environment. The focus is deliberately external and releasesafe: what the controller does to the simulation, not how the controller is built.

Across a baseline run and nine learned runs, the main behavioral signature is clear. In the no-control baseline, the solver first
enters runaway stress (RHS evaluations >= 100) at simulation time 50. In eight of nine learned runs, that onset is delayed to 72-
76, corresponding to a 44-52% shift. The median learned onset is 73 (+46% versus baseline). The longest sustained residence in the high-stress but pre-runaway band (80 <= RHS < 100) expands from 4 time units in the baseline to 9-12 in the strongest learned runs. A complementary dwell metric increases from 15.0 in the baseline to a learned median of 29.0.

The improvement is not monotonic. One run collapses back to the baseline onset. That matters. The correct claim is therefore not “perfect suppression” but recurrent threshold management: the neuromorphic controller repeatedly learns to postpone
runaway and hold the plasma edge near a quasi-stable boundary for materially longer intervals. That is already enough to justify a broader thesis. Neuromorphic control is valuable not because it imitates offline optimization, but because it can inhabit unstable physical regimes and reshape their timing.