Deep learning with artificial neural networks is increasingly gaining attention, because of its potential for data-driven astronomy. However, this methodology usually does not provide uncertainties and does not deal with incompleteness and noise in the training data. In this work, we design a neural network for high-resolution spectroscopic analysis using APOGEE data that mimics the methodology of standard spectroscopic analyses: stellar parameters are determined using the full wavelength range, but individual element abundances use censored portions of the spectrum. We train this network with a customized objective function that deals with incomplete and noisy training data and apply dropout variational inference to derive uncertainties on our predictions. We determine parameters and abundances for 18 individual elements at the ≈0.03dex level, even at low signal-to-noise ratio. We demonstrate that the uncertainties returned by our method are a realistic estimate of the precision and they automatically blow up when inputs or outputs outside of the training set are encountered, thus shielding users from unwanted extrapolation. By using standard deep-learning tools for GPU acceleration, our method is extremely fast, allowing analysis of the entire APOGEE data set in ten minutes on a single, low-cost GPU. We release the stellar parameters and 18 individual-element abundances with associated uncertainty for the entire APOGEE DR14 dataset. Simultaneously, we release astroNN, a well-tested, open-source python package developed for this work, but that is also designed to be a general package for deep learning in astronomy. astroNN is available at https://github.com/henrysky/astroNN with extensive documentation at http://astroNN.readthedocs.io.
This repository is to make sure all figures and results are reproducible by anyone easily for this paper [arXiv:1804.08622][ADS].
If Github has issue (or too slow) to load the Jupyter Notebooks, you can go http://nbviewer.jupyter.org/github/henrysky/astroNN_spectra_paper_figures/tree/master/
To get started, this paper uses astroNN developed by us and tested with astroNN 1.0.0 which Python 3.6 or above and reasonable computational resource is required. Extensive documentation of v1.0.0 at https://astronn.readthedocs.io/en/v1.0.0/ and quick start guide at https://astronn.readthedocs.io/en/v1.0.0/quick_start.html. astroNN Apogee DR14 Stellar Parameters and Abundances data is available as astroNN_apogee_dr14_catalog.fits.
Some notebooks make use of my milkyway_plot to plot on milkyway.
To continuum normalize arbitrary APOGEE spectrum with v1.0.0, see: https://astronn.readthedocs.io/en/v1.0.0/tools_apogee.html#continuum-normalization-of-apogee-spectra
- Datasets_Data_Reduction.ipynb
You should check out this notebook first as it describes how to reproduce the exactly same datasets used in the paper - Inference_highSNR.ipynb
It describes the inference process and result on spectra within SNR 100-200, also includes an isochrone plot and
jacobian analysis withastroNN_0617_run001. - Inference.ipynb
It describes the inference process and result on individual visit spectra which their combined counterpart in
training set withastroNN_0617_run001. - Inference_ApogeeBCNN.ipynb
It describes the inference process and result on spectra within SNR 100-200 withastroNN_0606_run001. - Open_Globular_Cluster_Benchmark.ipynb
It describes the inference on Open/Globular Cluster. - Apogee_dr14_NN_Catalog.ipynb
It describes how to generate stellar parameters and abundances for the whole APOGEE DR14, also contains plots of abundances across MilkyWay Galaxy. - ASPCAP_Normalization.ipynb
It describes how to compile dataset with ASPCAP normalized spectra (as opposed to continuum normalization), training and testing NN on that. - Small_Data.ipynb
It describes training neural network with smaller datasets and see the performance. - nn_figure6_draw_io
Source for Figure 6 in paper for the NN model, can be opened and edited by draw.io
astroNN_0606_run001is a trained astroNN's ApogeeBCNN() class model to infer 22 stellar parameters from APOGEE continuum normalized spectra.astroNN_0617_run001is a trained astroNN's ApogeeBCNNCensored() class model to infer 22 stellar parameters from APOGEE continuum normalized spectra.aspcapStar_BCNNCensoredis a trained astroNN's ApogeeBCNNCensored() class model to infer 22 stellar parameters from APOGEE ASPCAP-normalized spectra, with exactly the same model architecture asastroNN_0617_run001.small_data_fixed_****are trained astroNN's ApogeeBCNNCensored() class models with small dataset, with exactly the same model architecture asastroNN_0617_run001.
To load the model, open python outside astroNN_0606_run001 or astroNN_0617_run001
from astroNN.models import load_folder
# replace the name of the NN folder you want to open
neuralnet = load_folder('astroNN_0617_run001')
# neuralnet is an astroNN neural network object, to learn more;
# http://astronn.readthedocs.io/en/latest/neuralnets/basic_usage.html
# To get what the output neurones are representing
print(neuralnet.targetname)astroNN_apogee_dr14_catalog.fits is compiled prediction with astroNN_0617_run001 on the whole Apogee DR14. To load it with python
from astropy.io import fits
f = fits.getdata("astroNN_apogee_dr14_catalog.fits")
apogee_id = f['APOGEE_ID'] # APOGEE's apogee id
location_id = f['LOCATION_ID'] # APOGEE DR14 location id
ra = f['RA'] # J2000 RA
dec = f['DEC'] # J2000 DEC
# the order of the array is [Teff, log(g), C/H, C1/H, N/H, O/H, Na/H, Mg/H, Al/H, Si/H, P/H, S/H, K/H, Ca/H, Ti/H,
# Ti2/H, V/H, Cr/H, Mn/H, Fe/H, Co/H, Ni/H]
nn_prediction = f['astroNN'] # neural network prediction, contains -9999.
nn_uncertainty = f['astroNN_error'] # neural network uncertainty, contains -9999.To do inference on an arbitrary APOGEE spectrum,
- Open python under the repository folder but outside the folder
astroNN_0617_run001 - Copy and paste the following code to do inference with neural net in this paper on
2M19060637+4717296
from astropy.io import fits
from astroNN.apogee import visit_spectra, apogee_continuum
from astroNN.models import load_folder
# the same spectrum used in figure 5
opened_fits = fits.open(visit_spectra(dr=14, apogee='2M19060637+4717296'))
spectrum = opened_fits[1].data
spectrum_err = opened_fits[2].data
spectrum_bitmask = opened_fits[3].data
# using default continuum and bitmask values to continuum normalize
norm_spec, norm_spec_err = apogee_continuum(spectrum, spectrum_err,
bitmask=spectrum_bitmask, dr=14)
# load neural net
neuralnet = load_folder('astroNN_0617_run001')
# inference, if there are multiple visits, then you should use the globally
# weighted combined spectra (i.e. the second row)
pred, pred_err = neuralnet.test(norm_spec)
print(neuralnet.targetname) # output neurons representation
print(pred) # prediction
print(pred_err['total']) # prediction uncertaintyThePayne_dr14_catalog.fits is compiled from the data provided in the paper https://arxiv.org/abs/1804.01530
To load it with python
from astropy.io import fits
# the order is correspond to APOGEE DR14 allstar
f = fits.getdata("ThePayne_dr14_catalog.fits")
apogee_id = f['APOGEE_ID'] # APOGEE's apogee id
location_id = f['LOCATION_ID'] # APOGEE DR14 location id
ra = f['RA'] # J2000 RA
dec = f['DEC'] # J2000 DEC
# the order of the array is [Teff, log(g), C/H, C1/H, N/H, O/H, Na/H, Mg/H, Al/H, Si/H, P/H, S/H, K/H, Ca/H, Ti/H,
# Ti2/H, V/H, Cr/H, Mn/H, Fe/H, Co/H, Ni/H], same as astroNN DR14 order
payne_prediction = f['payne'] # ThePayne data, contains -9999.
# good flag is 1, bad flag is 0
payne_good_flag = f['good_flag'] # ThePayne quality flag- Henry Leung - henrysky
Student, Department of Astronomy and Astrophysics, University of Toronto
Contact Henry: henrysky.leung [at] utoronto.ca - Jo Bovy - jobovy
Professor, Department of Astronomy and Astrophysics, University of Toronto
The original header of the .txt file has been removed, the original header of the file is as follow:
Title: Calibrations of Atmospheric Parameters Obtained from
the First Year of SDSS-III Apogee Observations
Authors: Meszaros Sz., Holtzman J., Garcia Perez A.E., Allende Prieto C.,
Schiavon R.P., Basu S., Bizyaev D., Chaplin W.J., Chojnowski S.D.,
Cunha K., Elsworth Y., Epstein C., Frinchaboy P.M., Garcia R.A.,
Hearty F.R., Hekker S., Johnson J.A., Kallinger T., Koesterke L.,
Majewski S.R., Martell S.L., Nidever D., Pinsonneault M.H.,
O'Connell J., Shetrone M., Smith V.V., Wilson J.C., Zasowski G.
Table: Properties of Stars Used for Validation of ASPCAP
================================================================================
Byte-by-byte Description of file: aj485195t4_mrt.txt
--------------------------------------------------------------------------------
Bytes Format Units Label Explanations
--------------------------------------------------------------------------------
1- 18 A18 --- 2MASS The 2MASS identifier (1)
20- 27 A8 --- Cluster Cluster identifier
29- 35 F7.2 km/s RVel Heliocentric radial velocity
37- 42 F6.1 K Teff ASPCAP effective temperature
44- 49 F6.1 K TeffC Corrected ASPCAP effective temperature
51- 54 F4.2 [cm/s2] logg Log ASPCAP surface gravity
56- 60 F5.2 [cm/s2] loggC Log corrected ASPCAP surface gravity
62- 66 F5.2 [-] [M/H] ASPCAP metallicity
68- 72 F5.2 [-] [M/H]C ASPCAP corrected metallicity
74- 78 F5.2 [-] [C/M] ASPCAP carbon abundance
80- 84 F5.2 [-] [N/M] ASPCAP nitrogen abundance
86- 90 F5.2 [-] [a/M] ASPCAP {alpha} abundance
92- 97 F6.1 --- S/N Signal-to-noise
99-104 F6.3 mag Jmag 2MASS J band magnitude
106-111 F6.3 mag Hmag 2MASS H band magnitude
113-118 F6.3 mag Kmag 2MASS K_s_ band magnitude
120-124 F5.1 K e_TeffC The 1{sigma} error in TeffC
126-130 F5.3 [-] e_[M/H]C The 1{sigma} error in [M/H]C
--------------------------------------------------------------------------------
Note (1): After DR10 was published we discovered that four stars had double
entries with identical numbers in this table (those are deleted from
this table, thus providing 559 stars). All calibration equations were
derived with those four double entries in our tables, but because
DR10 is already published we decided not to change the fitting
equations in this paper. This problem does not affect the effective
temperature correction. The changes in the other fitting equations
are completely negligible and have no affect in any scientific
application. The parameters published in DR10 are off by <1 K in
case of the effective temperature error correction, and by < 0.001 dex
for the metallicity, metallicity error, and surface gravity
correction.
--------------------------------------------------------------------------------This project is licensed under the MIT License - see the LICENSE file for details