Shrinking the Haystack: One-Class Machine Learning Detection of Magnetosheath Current Sheets in MMS Burst Data
We present a morphology-first framework for narrowing the search for magnetic-reconnection candidates in Magnetospheric Multiscale (MMS) burst-mode data. The target is the small, short, and frequently electron-only reconnecting current sheets that occur in turbulent magnetosheath plasma. The pipeline operates in two stages. A local frame-quality gate based on minimum-variance analysis first retains only windows whose current-sheet coordinates are well defined. A one-class Deep Support Vector Data Description neural network then scores those windows against a library of 3,000 physically calibrated synthetic current sheets generated by Monte Carlo from a single published reference event. Acceptance into the surrogate library is governed by the second-order structure function $S_2(\tau)$: a candidate is admitted only if its multi-scale fingerprint tracks that of the seed event inside a tolerance band, together with a small number of shape-based checks. This $S_2(\tau)$-anchored construction defines the in-class distribution directly from a well-understood reference event and sidesteps the absence of a curated negative class in turbulent magnetosheath data. Applied to 15 magnetosheath turbulence intervals from the literature (1.58 h of burst-mode coverage), the framework compresses 22,775 sliding windows to 270 candidate detections (a 98.8% reduction). Manual visual screening identifies 93 of these as candidate reconnection events and a further 118 as sheet-like, retaining 78% of the queue for follow-up; the candidate-reconnection pool extends well beyond the 22 detections that overlap the published reconnection-event catalog used here as a sanity check. The framework is intended as the data-reduction stage of a broader reconnection-search workflow, offered here as an initial proof of concept before extending the one-class design to additional feature channels.