Training set
The v3 models (NRHJSur3dq8_AA_v3, NRHJSur3dq8_AdA_v3 and their shared
merger-ringdown component NRHJSur3dq8_MR_v3) were trained on 260 numerical
relativity simulations from the SXS catalogue. This page lists them.
Download
| file | contents |
|---|---|
| training-set-v3.json | the list with metadata (model generation, catalogue release, extrapolation order) and one record per simulation |
| training-set-v3.csv | one row per simulation: sxs_id,q,chi1z,chi2z,exchange_symmetric |
| training-set-v3.txt | the 260 SXS identifiers, one per line |
Each simulation can be retrieved with the sxs package, for example
sxs.load("SXS:BBH:2100", extrapolation="N2"). The identifiers carry no
version suffix; the models were built against SXS catalogue release v3.0.0.
Column definitions
sxs_id: the SXS catalogue identifier.q: the mass ratio, with body 1 the heavier, so thatq >= 1to within the catalogue's precision (the smallest listed value is 0.99985).chi1z,chi2z: the dimensionless spin components of bodies 1 and 2 along the orbital angular momentum.exchange_symmetric: 1 for the 18 simulations with|q - 1| < 0.02and|chi1z - chi2z| < 1e-3, and 0 otherwise. These binaries are invariant under exchange of the two bodies, so their odd-mmodes vanish; odd-mmismatches computed on them compare two numerically zero signals and are not meaningful.
The values of q, chi1z and chi2z are the catalogue's values at its
reference time, given to six decimal places.
Coverage
All 260 simulations are quasi-circular (catalogue eccentricity below 1e-3) and spin-aligned (in-plane spin components below 1e-3), with at least 18 orbits after the reference time. They span mass ratios from 1 to 8 and spin components up to 0.998 in magnitude.
Where several simulations share a parameter point (q, chi1z and chi2z
rounded to two decimal places), the longest was selected. Simulations found to
be under-resolved in a resolution-convergence study were replaced by the next
longest at the same parameter point; where no alternative existed, the
simulation was retained.
Of the 260 simulations, 151 lie within the parameter range of NRSur7dq4
(q <= 4, spin magnitudes <= 0.8).
Mirror pairs
Six pairs of simulations in the set are mirror images of each other: at
q = 1, one simulation is the other with the two spins exchanged, so each
pair describes the same binary with the bodies relabelled. The two members of
a pair are independent numerical realizations of the same physics, and their
difference is a model-independent estimate of the error of the training data.
| simulation | chi1z |
chi2z |
mirror | chi1z |
chi2z |
|---|---|---|---|---|---|
| SXS:BBH:4066 | -0.95 | +0.95 | SXS:BBH:4029 | +0.95 | -0.95 |
| SXS:BBH:3582 | -0.80 | -0.40 | SXS:BBH:2094 | -0.40 | -0.80 |
| SXS:BBH:3578 | -0.80 | +0.40 | SXS:BBH:2098 | +0.40 | -0.80 |
| SXS:BBH:2092 | -0.50 | +0.50 | SXS:BBH:0304 | +0.50 | -0.50 |
| SXS:BBH:2095 | -0.40 | +0.80 | SXS:BBH:3601 | +0.80 | -0.40 |
| SXS:BBH:2099 | +0.40 | +0.80 | SXS:BBH:3553 | +0.80 | +0.40 |
The pairs were identified from the list as simulations with |q - 1| < 0.01
whose spins match after exchange to within 0.01.
Origin of the list
The model files record the size of the training set but not the list of
simulations. NRHJSur3dq8_AA_v3.h5 records n_train = 260 and n_sym = 18
(the number of exchange-symmetric simulations) in its meta_json attribute,
and NRHJSur3dq8_AdA_v3.h5 records n_train = 260 as a root attribute. All
three v3 files record extrapolation = "N2". NRHJSur3dq8_MR_v3.h5 does not
record a training count; according to the build records it was trained on the
same 260 simulations, but this cannot be verified from the file itself.
The list was reconstructed on 2026-10-14 from two independent build records, which agree on the same 260 identifiers. Applying the selection criteria above to SXS catalogue release v3.0.0 on the same date yields the identical set. Because later catalogue releases may deprecate or add simulations, re-applying the selection criteria is not a substitute for this list.
Limitations
- This page covers the v3 model generation only.
- Accuracy measured against any of these simulations is in-sample for this model. Out-of-sample accuracy is given by the leave-one-out results, in which each simulation is predicted by a model trained on the remaining 259.