Galactic archaeology relies on millions of stellar spectra. To extract temperature, surface gravity, and elemental abundances from them, we need a fast, direct model. For decades, this has been replaced either by grid interpolation or by purely empirical “emulators” trained on precomputed results. Both approaches introduce their own errors: interpolation struggles with the high dimensionality of the element set, while emulators inherit the correlations of the training sample. Now that numerical simulation has been drastically accelerated, physics is once again the foundation of stellar analysis.
Payne Zero rewrites the synthesis logic: instead of sequentially stepping through wavelengths, they are all processed simultaneously on the GPU. This turns hours into seconds. For the atmospheric model, where iterations cannot be parallelized in the same way, critical sections are compiled for multi-core CPUs, and a trained initializer reduces the number of iterations. Atomic data calibration is performed via automatic differentiation of the entire radiative transfer chain, enabling simultaneous fitting of over one hundred thousand adjustments to oscillator strengths and damping parameters.
Synthesis of the solar spectrum on an H100 GPU took 14 seconds versus 760 seconds for the original code — a 54× speedup. For the APOGEE infrared range (1500–1700 nm), the time dropped to 1.4 seconds. Atmospheric iterations on 16 CPU threads became 6–7 times faster, and the trained initializer reduced the number of iterations by a further factor of 2–3. In a demonstration fit of 1600 real APOGEE stars, 12 elements (C, N, O, Mg, Al, Si, S, Ca, Ti, Mn, Ni) and the [Fe/H] parameter were determined simultaneously, with a physically self-consistent atmosphere recomputed at the end. The chemical sequences agree with the survey catalog, preserving the separation into high- and low-α-stars. Line calibration against atlases of the Sun and Arcturus reduced the discrepancy with observations by more than 80%.
Direct physical computation inside the optimizer eliminates the need for spectrum emulators. This means that any element can be introduced into the analysis without building new grids, and any discrepancies with the data can be examined at the level of atomic constants or atmospheric structure. Metallicity and individual abundances are now measured self-consistently, accounting for mutual influence through the equation of state and opacity. The method paves the way for mass recalibration of atomic data directly from observed spectra.
The next step will be to extend to cooler stars, where atmospheric iterations converge, and to models that deviate from LTE. It may be possible to create agent-based workflows in which different solvers are automatically selected based on the physical regime. Improved initializers that account for chemical groupings of elements will make direct fitting even faster.
The method will impact all fields that rely on large-scale spectral analysis: galactic archaeology, Milky Way dynamics, chemical evolution, and the characterization of exoplanet host stars.
It is necessary to test convergence on large samples of cool dwarfs and to improve the initializer for sparse abundance spaces. Integration with photometry and parallaxes is planned for joint determination of temperature and surface gravity.
The work directly addresses key problems in astrophysics: precise chemical tagging of stars to reconstruct the history of the galaxy, nucleosynthesis, and calibration of the distance scale. Fast direct synthesis reduces systematic errors in parameters that affect age and dynamics determinations.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This paper introduces Payne Zero, a direct and fast synthesis of stellar spectra on GPUs that extracts chemical abundances, temperature, and surface gravity from observations in seconds, without emulators or grid interpolation. The authors demonstrate a 54-fold speedup over the original code, a self-consistent fit of 12 elements for 1600 APOGEE stars, and automatic calibration of atomic data via differentiable radiative transfer.
Where its strength is
The main strength is placing the direct physical calculation inside the optimization loop, which removes the limitations of emulators and allows simultaneous refinement of both stellar parameters and atomic constants. This paves the way for large-scale spectroscopic analysis without compromising between speed and physical rigor.
Where it could go next
Worth trying
Why this may matter
If direct synthesis becomes the standard for large surveys, it will transform galactic archaeology: every spectrum will be analyzed with full physics, and systematics associated with emulators and interpolation will become a thing of the past. Mass recalibration of atomic data against observed spectra could improve abundance determinations for millions of stars, which is critical for understanding the chemical evolution of the Milky Way.
Nearby works in our archive
The closest work in the archive sits at 0.66 by meaning — the usual level for our corpus.
as written by the authors arXiv:2607.24141v1
Modern stellar surveys measure millions of spectra, yet one self-consistent atmosphere and spectrum can require tens of minutes. This cost has motivated grids, spectral emulators, and data-driven models. We present Payne Zero, which reorganizes one-dimensional LTE Kurucz calculations for GPU-native synthesis and multicore atmosphere iteration, and validate it against the original Fortran programs. A 300--1000 nm solar spectrum sampled at $$R_{\rm grid}=300{,}000$$ takes about 14 s on an NVIDIA H100 GPU, while the APOGEE 1500--1700 nm interval takes about 1 s. Physical atmosphere iterations take 2--5 s on 16 AMD CPU threads, and learned initializers reduce the iterations required for convergence. Final spectra remain in practical parity across the tested dwarf and giant regimes. These speeds place direct synthesis inside an optimizer without a label-to-flux spectral emulator. We demonstrate direct many-element fitting of reduced APOGEE spectra and recover multi-element abundance trends broadly consistent with the survey catalog. GPU-resident velocity shifts, broadening, line-spread-function convolution, and detector sampling add negligible cost relative to synthesis. The direct-synthesis search takes less than one minute per star on an H100, while atmosphere verification runs independently on multicore CPUs. The same computational graph calibrates more than $$10^5$$ oscillator-strength and damping corrections jointly against the Sun and Arcturus in about one minute on an H100. Payne Zero therefore brings direct physical fitting and atomic-data calibration to survey scale. The code is available at https://github.com/tingyuansen/payne-zero.
The project began as a hobby and made active use of large language models to translate code from Fortran to Python, but key decisions about data boundaries and parallelization remained with the human. Artificial intelligence translated the routine, but the architecture was guided by experience with the original programs.