Version 6 | Directional resolution in historical Brazilian aeromagnetic profiles
Run from the root of the unpacked supplement in the same Python environment used for v5 (NumPy, SciPy, pandas, Matplotlib).
  export OPENBLAS_NUM_THREADS=1
  python directional_resolution/sweep.py
  python directional_resolution/evaluate.py
  python directional_resolution/verify.py
  python directional_resolution/figure.py

These commands are offline and reuse the included distant_transfer and matched_rotation inputs and archived scores. PROTOCOL.txt records the order of the analyses and the follow-up with disjoint training disks.

Geometry and units. The hypothetical 1971 field is expressed in the tangent frame of the real 2008 field. A bearing of zero points toward increasing downward inclination, and +90 degrees points toward increasing eastward declination. The forward change in geographic declination is a different quantity. In the manuscript’s earlier 15-degree cone calculation, angles were offsets from each true historical tangent, measured about the 2008 vector. The angles in this directory are absolute bearings in the v5 tangent convention. The predicted field is at the historical flight height, with the separate 2008 reference at 2.2 km. Each historical segment retains the true angular separation and intensity ratio in that tile at those elevations. No new 2008 source coefficients are estimated from historical observations. A calibration step using historical observations chooses only a single common bearing per training sample. All errors are point-weighted nT.

Efficient scoring. For each observation, the scalar prediction with aligned moment and projection directions is a quadratic form in the unit vector on the cone. At a fixed opening angle it is a second-degree trigonometric polynomial of bearing. Segment centering and projection off a linear east-coordinate trend are linear operations. Each sum of squared errors is therefore a fourth-degree trigonometric polynomial. Nine equally spaced angular samples determine it exactly to floating-point precision. Each segment is checked against direct evaluation at the true bearing, at the representative offshore bearing and at 113.6 degrees, which lies outside the interpolation sample. The last direction is also checked against the original dipole kernel independently of the tensor implementation. Both metrics are checked against v5 scores for the real field and the marine comparison. The true local benchmark varies by tile and flight height and need not lie on the integer-degree grid.

Outputs. segment_coefficients.csv.gz contains each segment's region, row ID, production line, 12 km tile indices, 24 km block, count, bearing, opening angle, baseline squared errors and Fourier coefficients. In order these coefficients are the constant, cos(k theta) for k=1..4 and sin(k theta) for k=1..4. cone_curves.csv evaluates the total curve on all 360 grid bearings. cone_summary.csv records the optimum, benchmark, min/max 3D vector difference from the historical field at the optimum, number of grid bearings as good as the historical benchmark, and the inclusive contiguous integer range within 5% of the best RMSE. This descriptive width depends on the baseline error, sampling and criterion; profiles have different baseline errors. It is not a confidence interval. tolerance_widths.csv gives inclusive counts and ranges at four relative and three absolute tolerances. At 0.10 nT, full-shape ranges comprise 37 nearby versus 68 distant bearings. The curvature ranges both comprise 63.

withheld_folds.csv identifies the geographic block (or the entire opposite region), numbers of observations and segments, selected bearing and held-out SSE. spatial_line excludes shared production-line IDs, the full test block and all calibration centers within 24 km of any test tile center. disjoint_disks additionally excludes center distance <=48 km; nearby blocks -1_1 and 0_1 have no eligible calibration segment and are omitted. The nearby disjoint result uses 486 points across seven blocks. cross_region trains on the opposite region, excluding shared line IDs and any center within 48 km of a target center; all source segments remained eligible. withheld_summary.csv compares models on exactly the test points scored in each row. verify.py reconstructs each fold independently and checks the stated cutoffs, rankings and summary. Figure_7.png uses different y-axis limits in the two panels. These are exploratory comparisons on Brazilian surveys, not independent replication in another survey or a reconstruction of the physical geomagnetic field from the profiles alone.
