Two-channel diagnostics for charged-particle integrators: reproduction bundle (v1.0.0)
Year: 2026
Type: software
Abstract: Reproduction bundle for the study of how a readout floor can hide a secular growth of the energy error from an absolute-level diagnostic. It contains the experiment scripts that produce every measured number in the manuscript, the committed JSON and NPZ outputs of those scripts, the four figure scripts, the trained defect-corrector checkpoints and a single-command reproduction driver (code/repro.py). The physical setting is a charged particle in a non-stationary magnetic field, integrated by the Boris scheme, by a learned defect corrector with a hard speed-constraint projection, by three external machine-learning architectures (Hamiltonian neural network, symplectic network, symplectic physics-informed network) and by five classical schemes: a staggered Boris variant, two volume-preserving splittings of second and fourth order, the implicit midpoint rule and a two-stage Gauss-Legendre method. The analysis separates the magnitude and direction channels of the velocity error, measures the readout floor of the energy error, scores every scheme against a closed-form reference rather than against a refined run of the scheme under test, and reports an applicability map over 120 step-and-field cells, twenty independent retrainings, a hyper-parameter search over 96 configurations and seven ablations of the training objective. The result is in part negative, and this bundle is the evidence for it. The learned corrector is outperformed by the fourth-order volume-preserving splitting at the same step and at a fraction of its cost, and its trajectory advantage inverts beyond roughly one hundred gyro-orbits. Every number supporting that statement is in the committed outputs, and the scripts that produced them are here. The accompanying manuscript has not yet been submitted; no journal is named in this record on purpose. Licence: MIT. See CITATION.cff for citation metadata and code/README.md for how to run the bundle.
Language: English
Subfield: Numerical Analysis
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7207601022