Physical spectra can again be calculated during fitting
Stellar spectroscopy moved from repeatedly operated physical programs to precomputed grids and learned spectral emulators because atmosphere and synthesis calculations were too slow to evaluate for every trial. Emulators made high-dimensional fitting practical, but they also placed an interpolated label-to-flux relation between the requested stellar state and the physical calculation that generated the training library.
Payne Zero reorganizes the one-dimensional, plane-parallel, local thermodynamic equilibrium (LTE) calculation for multicore CPUs and accelerator-parallel synthesis. In the demonstrated regime, the physical model can therefore remain inside spectrum generation, fitting, and line-data calibration. This does not remove missing physics, inaccurate line data, or instrument mismatch. It makes those assumptions and residuals direct parts of the calculation that can be inspected and improved.
The name “Zero” refers to zero learned spectral-emulation layers. Every trial flux is calculated from populations, opacity, line profiles, and radiative transfer. A learned model is still used where it has a narrower role: predicting a useful starting atmospheric structure. Physical atmosphere iterations, not that prediction, establish the accepted converged state.
What the project provides
Payne Zero calculates one-dimensional, plane-parallel LTE stellar atmospheres and emergent spectra. Atmosphere correction runs as an ordered multicore CPU calculation. Spectrum synthesis evaluates independent wavelength work with PyTorch on CUDA, Apple Metal, or CPU. The same synthesis engine is used for model generation, observed-spectrum fitting, and line-data calibration.
The public source, issue tracker, installation scripts, tutorial notebook, and package documentation are in the Payne Zero GitHub repository. The homepage summarizes the paper's Kurucz comparison and representative runtime measurements.
Citation
If Payne Zero contributes to a publication, cite Ting and Kim, The Payne Zero Project I: Stellar Spectra from Physical Models in Seconds.
@unpublished{TingKim2026PayneZero,
author = {Ting, Yuan-Sen and Kim, Elliot M.},
title = {The Payne Zero Project I: Stellar Spectra from Physical Models in Seconds},
year = {2026},
note = {Manuscript in preparation}
}This temporary entry reproduces the manuscript title without inventing a preprint identifier. It will be replaced by the public bibliographic record after release.
License, sources, and scientific lineage
Payne Zero-authored code is released under the BSD 3-Clause License. The source distribution includes its license, notices, and manifests for runtime data and generated assets. Catalog and source-table versions remain part of the scientific record.
The physical framework descends from Robert L. Kurucz’s ATLAS and SYNTHE programs. Payne Zero is a separate package rather than an update or renamed release of those programs. It rewrites substantial parts of the implementation and defines its own packages, interfaces, structured products, hardware execution, fitting workflow, and line-calibration system.
This website uses the paper for scientific context and the repository, README, tutorial notebook, and public source for runnable interfaces. If a manuscript description and the released interface differ, the checked source code defines what the installed package can run.