Tomasz Różański, Yuan-Sen Ting and Maja Jabłońska, "Toward a Spectral Foundation Model: An Attention-Based Approach with Domain-Inspired Fine-Tuning and Wavelength Parameterization", https://arxiv.org/abs/2306.15703
Important note: The implementation of TransformerPayne in this repository includes slight modifications from the version described in the paper above. These changes involve most importantly the placement of residual connections and hyperparameter tuning. Detailed explanations will be provided in an upcoming paper.
import jax
jax.config.update("jax_enable_x64", True)
import transformer_payne as tp
import numpy as np
import matplotlib.pyplot as plt
emulator = tp.TransformerPayne.download() # Download weights for default emulator
wave = np.linspace(4670, 4960, 20000)
mu = 1.0 # Ray perpendicular to solar surface
parameters = emulator.solar_parameters # Pick solar parameters and abundances
spectrum = emulator(np.log10(wave), mu, parameters) # Emulate a spectrum
Visulize the results:
intensity = spectrum[:,0]
continuum = spectrum[:,1]
normalized_intensity = intensity / continuum
fig, axs = plt.subplots(2, sharex=True, figsize=(5, 4))
axs[0].plot(wave, intensity, color='black', label='Intensity')
axs[0].plot(wave, continuum, color='red', label='Intensity continuum')
axs[1].plot(wave, normalized_intensity, label="Normalized intensity", color='black')
axs[1].vlines(4862.721, 0, 1, color='C0', label=r'$H_\beta$ (vacuum wavelength)') # Using raw string for LaTeX
axs[1].set_xlabel("Wavelength [$\AA$]") # Corrected to set xlabel on the second subplot
axs[0].set_ylabel("Intensity [erg/s/cm$^3$/ster]")
axs[1].set_ylabel("Normalized Intensity")
axs[0].legend(loc="lower left")
axs[1].legend(loc="lower left")
plt.subplots_adjust(hspace=0.05)
plt.show()
Install transformer-payne along with the huggingface-hub and joblib modules, which enable easy downloading of model weights from the internet.
Make sure you also have matplotlib and numpy installed in order to run the example shown above.
If you plan to use the notebooks with additional examples located in the tutorial/ directory, please also install jupyterlab, notebook, and ipykernel.
pip install transformer-payne
pip install huggingface-hub joblib numpy matplotlib
pip install jupyterlab notebook ipykernel
git clone https://github.com/RozanskiT/transformer_payne.git
cd transformer_payne
pip install -e .[dev]
Disclaimer: This is an experimental version of the emulator. We do not recommend using it to infer atmospheric parameters of stars, as it currently has significant limitations. However, it is valuable for method development and benchmarking. We are actively working on a future release of emulators with well-understood and validated precision suitable for inference. If you have any questions, concerns, or are interested in developing an emulator for your own spectral grid (in intensity, fluxes, or potentially including Stokes parameters), please don’t hesitate to contact me.
This document outlines an emulator built upon a slightly adapted version of the Korg.jl package and the MARCS stellar atmosphere grid. It simulates stellar spectra across a broad parameter space:
- Effective Temperature (Teff): 4,000 - 8,000 K
- Surface Gravity (logg): 2.0 - 5.0
- Microturbulence (vmic): 0 - 5 km/s
- Metallicity ([Fe/H]): -2.5 to 1.0
- Alpha-element Enhancement ([alpha/Fe], ): -1.0 to 1.0
- Carbon-to-Iron Ratio ([C/Fe]): -1.0 to 1.0
- Elemental Abundances: Individual abundances can vary within a logarithmic range of ±0.3, from Helium to Uranium (with respect to given [Fe/H], [alpha/Fe] and [C/Fe])
- Resolution: ~300,000
- Wavelengths span: 1,500 to 20,000 angstroms
The emulator employs the Kurucz's linelist gfall08oct17.dat in Korg.jl code. For more information, visit the following links: