Based on the papers by Hon et al. (2017) and Hon et al. (2018). The deep learning classifier learns to classify folded frequency power spectra of red giants as either hydrogen shell-burning (RGB; label 0) or core helium-burning (HeB; label 1).
The classifier requires background-subtracted frequency power spectrum of a red giant, along with measurements of delta_nu (large frequency separation) and numax (frequency at maximum oscillation power). It is trained using the consensus asteroseismic evolutionary states from APOKASC (Elsworth et al. 2019), which has labels for red giants down to a delta_nu of 2.8uHz.
- numpy
- scipy
- pandas
- torch (>= 0.4.0)
To perform inference on a star, download the folder and run inference.py. The script accepts the following arguments:
- '--classifier': Path to classifier. Default is 4-year classifier in the /saved_models directory. A 3-month classifier is also provided
- '--psd_file': Path to ASCII file of backgrond-corrected power spectrum. Needs to have frequency in first column and power in the second. Can be either whitespace or comma separated.
- '--numax': Frequency at maximum oscillation power in uHz
- '--dnu': Large frequency separation in uHz
- '--mc_samples': Number of Monte Carlo samples for uncertainty estimation
- '--plot_spectrum': Boolean flag specifying whether the spectrum should be plotted in tandem with prediction
python inference.py --psd_file example_ps/1027337.csv --numax 74.21 --dnu 6.937
python inference.py --psd_file example_ps/2014377.dat --numax 39 --dnu 4.046
The delta_nu and numax values in these examples are taken from the measurements of 16,000 Kepler red giants by the SYD pipeline.
The deep learning algorithm is a Bayesian convolutional neural network that is based on Bayes by Backprop. It borrows heavily from the implementation in this repo.
This repo will be updated soon with scripts to train a network from scratch.