Author ORCID Identifier
https://orcid.org/0009-0006-6520-4495
Date of Award
Summer 6-8-2026
Document Type
Thesis (Undergraduate)
Department
Physics and Astronomy
First Advisor
Robyn M. Millan
Second Advisor
Ryan C. Hickox
Third Advisor
James W. LaBelle
Abstract
The open–closed magnetic field line boundary (OCB) demarcates closed terrestrial field lines from those threaded into the interplanetary medium, and its latitude encodes the instantaneous balance between dayside and nightside reconnection that governs much of the high-latitude space-weather hazard. No single instrument resolves it globally in real time. In this thesis I close that gap empirically. Pairing nearly three decades (1983–2012) of Defense Meteorological Satellite Program particle-precipitation boundaries with the 5-minute OMNI solar-wind and geomagnetic-index record yields ∼9.1×105 causally matched crossings, partitioned into six hemispheric/MLT sectors after a |MLat| ≥ 40° physics cut.
On this corpus I train three complementary predictors. A Ridge + RandomForest + XGBoost stacking ensemble achieves a best-sector test RMSE of 2.54° magnetic latitude on dayside_north and outperforms Newell-coupling, persistence, and climatology baselines in every sector. A median-split binary classifier reaches 80% accuracy, but best-sector balanced accuracy collapses to 0.51 and 0.26 at 4- and 10-bin equal-frequency targets respectively (sector-median 0.48 and 0.24)—evidence that the OCB at this cadence is a continuous boundary above a ∼2–3° apparently irreducible noise floor, not a finite-state latitude system. A multi-quantile XGBoost regressor with split-conformal calibration restores empirical 80% coverage to within ±0.02 of nominal across all six sectors after a small ε shift of 0.17–0.85°, supplying calibrated probabilistic OCB rings rather than overconfident point estimates.
Embedding magnetospheric physics directly in the learner sharpens both skill and interpretability. A residual-XGBoost model trained atop a frozen Newell-only baseline becomes the overall winner, taking 5 of 6 sectors; a monotone-constrained XGBoost matches it without measurable RMSE penalty. A physics-informed Keras MLP whose loss penalizes departure from a discretized expanding/contracting polar-cap flux-balance equation, posed in magnetic colatitude, recovers hemisphere-invariant coefficients α ∈ [+1.23, +1.83] (a consistently positive Newell-driven source efficiency) and β ≈ 0 (no resolvable AE-indexed sink at 5-minute cadence) across every sector. SHAP attribution exposes a clean dayside/nightside dichotomy, a quantitative fingerprint of the Dungey cycle, with solar-wind drivers and geomagnetic indices carrying 86–89% of the budget on dayside and northern sectors and cyclical UT contributing 36–38% on the southern nightside; a model-free test, however, traces that strongest hemispheric signal mainly to the larger southern geomagnetic-pole offset, isolating a genuine physical effect from the DMSP sampling bias with which it is otherwise confounded. Appending the most recent boundary as an autoregressive predictor lowers RMSE by up to 0.35° on the nightside sectors (largest for linear models)—an empirical echo of the expanding/contracting polar-cap paradigm—while an auxiliary precipitation-boundary prior helps tree models but not linear ones. Together, these results establish a fast, physics-consistent, and calibrated baseline for empirical OCB specification, and quantify exactly which observational gaps a future multi-satellite constellation would have to close.
Recommended Citation
Singh, Arnav, "TRACING THE SEAM: MACHINE LEARNING MODELS OF THE OPEN–CLOSED BOUNDARY FROM UPSTREAM SOLAR WIND DRIVERS" (2026). Physics and Astronomy Undergraduate Senior Theses. 4.
https://digitalcommons.dartmouth.edu/physics_senior_theses/4
