Stellar spectra from
physical models,
in seconds.

Stellar spectra respond to temperature, gravity, chemical composition, and atmospheric structure together. A change in one element can alter its own lines, blended features, molecular equilibrium, the supply of free electrons, continuum opacity, and sometimes the atmosphere itself. A physical forward model keeps those connections inside the calculation.

Spectral emulators made repeated stellar-parameter fitting practical when atmosphere and synthesis programs were too slow to call for every trial. Payne Zero reorganizes the same physical target for modern multicore CPUs and accelerators. “Zero” means that no learned label-to-flux layer stands between the requested stellar state and its spectrum: opacity and radiative transfer are evaluated directly. A learned model is still useful for predicting the starting atmosphere, while the physical solver decides whether a retained atmosphere has converged.

Begin by choosing how to describe the chemical mixture. Payne Zero uses that same description to initialize the atmosphere in either workflow. You can then synthesize while holding the initialized structure fixed, or iterate the atmosphere to convergence before synthesizing. The abundance choice and the atmosphere treatment are separate decisions.

Bracket notation measures an abundance relative to the Sun on a logarithmic scale: for example, [Fe/H] = −0.5 means one third of the solar iron-to-hydrogen ratio. A difference of 1 dex is a factor of ten. “Alpha enhancement” describes the shared behavior of elements such as O, Mg, Si, S, Ca, and Ti relative to the bulk metal content.

payne-zero-atmosphere \
  --effective-temperature 5777 \
  --log-surface-gravity 4.44 \
  --metallicity 0.0 \
  --alpha-enhancement 0.0 \
  --microturbulence-km-s 1.0 \
  --out runs/sun_atmosphere

payne-zero-synthesis \
  runs/sun_atmosphere/payne_zero_structured_atmosphere.npz \
  --wl-start-nm 500 \
  --wl-end-nm 510 \
  --r-grid 300000 \
  --out runs/sun_converged_spectrum.npz

from payne_zero_atmosphere import solve_structured_atmosphere
from payne_zero_synthesis import synthesize

solar_labels = dict(
    effective_temperature=5777,
    log_surface_gravity=4.44,
    metallicity=0.0,
    alpha_enhancement=0.0,
    microturbulence_km_s=1.0,
)

atmosphere_path = solve_structured_atmosphere(
    **solar_labels,
    out_dir="runs/sun_atmosphere",
)

spectrum = synthesize(
    atmosphere_path,
    wavelength_start_nm=500,
    wavelength_end_nm=510,
    resolution=300_000,
    device="auto",
)

spectrum.save_npz("runs/sun_converged_spectrum.npz")

payne-zero-synthesis \
  --effective-temperature 5777 \
  --log-surface-gravity 4.44 \
  --metallicity 0.0 \
  --alpha-enhancement 0.0 \
  --microturbulence-km-s 1.0 \
  --wl-start-nm 500 \
  --wl-end-nm 510 \
  --r-grid 300000 \
  --out runs/sun_initialized_spectrum.npz

from payne_zero_synthesis import synthesize_from_labels

spectrum = synthesize_from_labels(
    effective_temperature=5777,
    log_surface_gravity=4.44,
    metallicity=0.0,
    alpha_enhancement=0.0,
    microturbulence_km_s=1.0,
    wavelength_start_nm=500,
    wavelength_end_nm=510,
    r_grid=300_000,
    device="auto",
)

spectrum.save_npz("runs/sun_initialized_spectrum.npz")

For four matched stellar controls over 300–1000 nm at R = 300,000, the pooled RMS difference from the Kurucz calculation is below one part in a thousand in normalized flux. This remains true when Payne Zero and the original programs first solve their atmospheres independently. This is a numerical implementation comparison, not a claim of agreement with every observed stellar spectrum.

In the paper's warm repeated-call measurement, the 300–1000 nm solar spectrum took 14 seconds on one NVIDIA H100. Over the narrower APOGEE 1500–1700 nm interval, synthesizing a solar spectrum from the same atmosphere took 1.4 seconds on the H100 at R = 300,000, before instrument convolution or resampling. One physical atmosphere pass took 2.1–5.3 seconds on 16 AMD EPYC threads across the four controls. These timings exclude first-use compilation and depend on wavelength range, R, device, numerical type, and cache state.