Skip to content

Secondary-photon fit form quality across 21 materials

This is a regression-testing harness rather than an analytic verification script. It exists to answer the question "is my coupled-photon B_p(τ) fit form good enough across the materials this library is likely to be applied to?", and is the tool that was used to design the shipped Form K_reg fit.

Setup

The cache file verification_and_validation/secondary_photon_fit_cache.json holds 6 training anchors plus 1 held-out anchor for each of 21 shielding materials, generated with compute_buildup against ENDF/B-VIII.1 with weight windows on. Materials span:

  • Hydrogenous moderators: polyethylene, polyethylene_deep, borated_poly, lih, d2o, h2o_water (pure water).
  • Mixed light/heavy composites: WC + 20/50/70 vol% water, WB + 20/50/70 vol% water (peak-and-decay \(B_p\)).
  • Saturating heavy attenuators: concrete, steel, ss316, tungsten, lead, magnetite (heavy concrete), b4c, aluminum, concrete_steel10.

Per material, the held-out anchor sits ~30-50% past the deepest training anchor.

What the harness reports

For each candidate fit form in CANDIDATES, for every material:

  • Training fit z-score — max |B_pred - B_mc| / σ over the 6 training anchors. Catches in-sample over-/under-fitting.
  • Holdout z-score — |B_pred - B_mc| / σ at the held-out anchor. Catches over-fitting that hides in training but shows up in extrapolation.

Aggregated across the 21 materials:

Current shipped form (Power × sat, 3p) Form K_reg (5p, L2)
Worst holdout z 54.3 36.6
Mean holdout z 14.7 6.0
Median holdout z 7.1 2.4
Worst training z 42.4 3.3

K_reg dominates baseline on every metric. The wins come from the late-stage exponential growth tail (exp(d·x_n + e·x_n²) with signed d, e) which baseline's pure power-law saturation cannot represent.

Running

python verification_and_validation/secondary_photon_fit_test_suite.py

The script also writes:

  • secondary_photon_fit_dataset.csv — flat (material, τ, descriptors, B_p, σ) CSV ready for symbolic regression or external analysis (gitignored).
  • /tmp/secondary_photon_fit_suite_plot.png — per-material comparison panels.

Adding a new candidate form

Drop a (fit_fn, eval_fn) pair into the CANDIDATES dict in secondary_photon_fit_test_suite.py. Re-run; the harness scores it across all 21 materials and prints the per-material residuals plus the worst / mean / median holdout z.

Same machinery for flux

The harness loads the dose-AP-coupled-photon field per material. The secondary-photon flux is fit with the same Form K_reg in the shipped library; to validate that path:

  1. Regenerate the cache including the flux quantity: compute_buildup(quantities=["dose-AP-coupled-photon", "flux-coupled-photon"], ...).
  2. In _load(), swap the read of r["mc"]["dose-AP-coupled-photon"] for r["mc"]["flux-coupled-photon"] (and the matching pk, mc_std_dev).

Same applies for the neutron and primary-photon sibling harnesses.