mjablonska/TPayne-spice-harps-v2

TPayne-spice-harps v2 (corrected C/N/O grid)

TransformerPayne emulator of MARCS + Turbospectrum specific intensities over the HARPS range (3780–6910 Å), for use with SPICE (stellar-spice). Release maja-harps-cno-v2-intensity-tpayne-12p-2ch-192x6-500k@0.3.1.

This bundle replaces the first release (RozanskiT/TPayne-spice-harps, revision 3df764e). Two changes matter:

• Corrected C/N/O. The first grid passed the [C/Fe], [N/Fe] and [O/Fe] offsets to Turbospectrum as absolute abundances, so its spectra were synthesised with essentially no carbon, nitrogen or oxygen (no CH, CN, C₂ or CO). The v2 grid uses the corrected interface. This was checked against Korg, five Gaia FGK benchmark stars, the solar flux atlas and AMBRE.
• 12 inputs. marcs_fe_h, the metallicity of the MARCS structure, is now an input. SPICE fills it with feh by default (marcs_fe_h_mode="feh"); "snap" uses the nearest MARCS node instead.

Inputs

name	meaning	range
marcs_teff	Teff of the nearest MARCS atmosphere (K)	3000–8000
marcs_logg	log g of the nearest MARCS atmosphere	−0.5–5.5
feh	[Fe/H] of the line synthesis	−5–+1
marcs_fe_h	[Fe/H] of the atmospheric structure	−5–+1
vmicro	microturbulence (km/s)	1–5
a	[α/Fe] offset	−0.4–+0.8
c, n, o	[C/Fe], [N/Fe], [O/Fe] relative to solar-scaled (Magg et al. 2022)	see input_domain
r, s	r- and s-process scalings	−0.4/−0.5–+1
mu	cosine of the viewing angle	0.01–1

Outputs: min–max scaled log10 line and continuum intensities (two channels). Wavelengths are evaluated on any grid inside the range.

Usage

Loading the bundle needs astro_emulators_toolkit ≥ 0.2.0. Version 0.1.0 cannot read it ("Unknown training config keys").

from spice.spectrum.aemu_spectrum_emulator import IntensityPretrainedAemuSpectrumEmulator
emu = IntensityPretrainedAemuSpectrumEmulator("mjablonska/TPayne-spice-harps-v2")
params = emu.to_parameters(dict(marcs_teff=5750, marcs_logg=4.5, feh=0.0, vmicro=1.0))  # marcs_fe_h <- feh

Training data

• 235,971 Turbospectrum (v19, LTE, GES v5 line lists) intensity spectra on MARCS atmospheres at 0.01 Å sampling. The spectra are not convolved.
• The rows are 185,975 Latin-hypercube rows plus 49,996 targeted at cool, metal-rich and low-gravity stars. Teff and log g take MARCS node values only; [Fe/H], vmicro, the abundance offsets and μ are continuous.
• Molecular line lists start at 4200 Å, so 3780–4200 Å contains atomic lines only.
• 10 rows of very cool, metal-rich giants are missing: Turbospectrum's molecular equilibrium does not converge there.

Training recipe.

• Architecture: TransformerPayne, 192 dimensions × 6 layers, 8 tokens.
• Optimisation: SOAP with lr 1e-3, warm-up 50k steps, cosine schedule; 500k steps; gradient clipping at 1.0; TF32 matmuls.
• Data: 2048 wavelengths per row, 64 rows per batch.

Validation

4,096 held-out spectra. Normalised-intensity residuals:

metric	value
median per-spectrum rms	0.0060
rms over all spectra	0.0095
fraction of the rms on scales below 2 Å	87%
pixels off by more than 1%	13%
pixels off by more than 2%	4.3%
relative MAE, intensity	1.10%
relative MAE, continuum	0.27%

• Targeted rows (not seen in training by this or the previous v2 bundle): v2 has rms 0.0098; this release has 0.0088.
• Off-node. On Turbospectrum spectra computed on atmospheres interpolated between MARCS nodes, the median rms is 0.0069; at the nodes it is 0.0080.

Known limitations:

• In the Hβ wings of hot dwarfs (8000 K, log g 4) the residual is 1.2–1.4% rms at μ ≥ 0.7, with spikes up to 7%.
• Errors grow towards cool, metal-rich stars: median rms 0.013 at [Fe/H] > +0.5.
• Abundances are on the Magg et al. (2022) scale. A solar-composition star fitted in 1D LTE returns [C/Fe] ≈ −0.17.

Training history: W&B run twhtojgy (project tpayne-spice-harps).