TPayne-spice-harps-interp

A Transformer Payne emulator of the emergent specific intensity I(lambda, mu) of FGKM stars over the HARPS range (3780-6910 A), trained on Turbospectrum spectra computed with interpolated MARCS model atmospheres. It is the intensity emulator used by SPICE (https://arxiv.org/abs/2511.10998) for spectra and photometry of inhomogeneous stellar surfaces.

Compared with mjablonska/TPayne-spice-harps-v2 (12 inputs, trained on spectra computed with the nearest MARCS node), this model takes the stellar parameters directly: there are no marcs_* inputs, and its predictions have no steps where the old grid switched atmosphere node (e.g. 30-60 mmag in B at [Fe/H] node midpoints and at microturbulence 3.5 km/s in the v2b model; here below 0.01 mmag along fine one-parameter scans).

Inputs (11) and outputs

input	meaning	training range
teff	effective temperature (K)	3000-8000
logg	surface gravity log g (cgs)	-0.5 to 5.25
feh	[Fe/H] of line formation and atmosphere	-5 to +1
vmicro	microturbulence of line formation (km/s); the atmospheres use the 2 km/s MARCS models	1-5
a	[alpha/Fe]	-0.4 to 0.8
c	[C/Fe]	-1 to 1
n	[N/Fe]	-0.5 to 1
o	oxygen relative to the alpha enhancement: A(O) = 8.77 + [Fe/H] + a + o, i.e. [O/Fe] = a + o	-0.4 to 0.8
r	[r/Fe] (Turbospectrum R-PROCESS)	-0.4 to 1
s	[s/Fe] (Turbospectrum S-PROCESS)	-0.5 to 1
mu	cosine of the viewing angle	0.01-1

Outputs: two channels at any wavelength in 3780-6910 A (input as log10 Angstrom): log10 of the specific intensity with lines and log10 of the continuum intensity. Spectra are not convolved with any instrumental or rotational profile.

Training

• Data: the v2 release of mjablonska/harps-interp-t02-grid (299,986 spectra; LTE Turbospectrum, interpolated MARCS atmospheres, spherical below log g 3, solar abundances of Magg et al. 2022; molecular CN and CH lines included below 4200 A): 295,817 training and 4,096 validation spectra; the 73 rows with non-finite continuum pixels (all below 3200 K) were excluded.
• Model: Transformer Payne, 192 dim, 6 layers, 8 tokens; SOAP optimiser (lr 3e-4, cosine schedule, gradient clipping 1); batch 64 spectra x 2048 wavelengths sampled on the native 0.01 A pixels.
• Schedule: 500k steps (cosine) + warm restart to 1M steps on the v1 grid (no molecular lines below 4200 A), then a 250k-step warm-restart fine-tune on the v2 grid (1.25M steps; this was v1.0.0).
• Balmer-wing smoothness fine-tune (v1.1.0): 1.55M further steps from v1.0.0 (300k + 250k + 1M, three warm restarts, peak lr 1e-4 / 5e-5 / 5e-5, cosine) on a 133,120-spectrum subset of the training set. For every star with Teff >= 6500 K, 512 of its 2048 pixels per batch are a contiguous 5 A run inside an H-alpha to H-delta wing (3 A from the core out to 40-60 A), and the loss adds (weight 1) the mean square of the emulator-minus-Turbospectrum residual on that run after removing a quadratic in wavelength. The penalty acts on the residual, so real metal lines on the wings are fitted, not smoothed.
• Validation: 4096 held-out spectra at fixed pixels; final MSE 3.6e-7 in min-max scaled log10 intensity (v1.0.0: 3.8e-7).

Use in SPICE

from spice.spectrum.aemu_spectrum_emulator import IntensityPretrainedAemuSpectrumEmulator

emu = IntensityPretrainedAemuSpectrumEmulator("mjablonska/TPayne-spice-harps-interp")
params = emu.to_parameters({"teff": 5750, "logg": 4.4, "feh": 0.0, "vmicro": 1.0,
                            "a": 0.0, "c": 0.0, "n": 0.0, "o": 0.0, "r": 0.0, "s": 0.0})

Needs JAX_ENABLE_X64=1 (set it before importing JAX) and astro-emulators-toolkit 0.2.0.dev0 or newer: older toolkits reject the bundle config (Unknown training config keys). Tag v1.1.0 marks this release (Balmer-wing fine-tune). The previous model is tag v1.0.0: download it with huggingface_hub.snapshot_download("mjablonska/TPayne-spice-harps-interp", revision="v1.0.0") and pass the returned local directory instead of the repo id.

Known limitations

• The input box is not filled uniformly: points without a complete MARCS bracket (e.g. hot low-gravity stars near the Eddington limit) are absent from the training grid, so predictions there are extrapolations.
• Below 4200 A only CN (12C14N) and CH (12CH, 13CH) molecular lines are included (no 13CN, C15N, NH or OH there).
• Hot stars: v1.0.0 and earlier emulators put about 1% of spurious small-scale structure (~1 A ripple) on the H-beta wings at 8000 K. The v1.1.0 fine-tune reduces it to 0.3% (8000 K, log g 4.0, mu = 1; the emulator's invented structure on flat H-beta wing pixels of hot validation stars is 0.13%, as in metal-line regions). It is a fixed pattern in wavelength, so it does not average down over many spectra of similar stars.
• LTE; one-dimensional hydrostatic atmospheres.

Provenance

v1.1.0: training run smooth_wings/ft_pen1y (bundle release name maja-harps-interp-sw-pen@0.0.1), fine-tuned from v1.0.0 = release maja-harps-interp-t02v2-intensity-tpayne-11p-192x6-1m25@0.3.0; astro-emulators-toolkit 0.2.0.dev0. Contact: Maja Jablonska.