Source-linked AI summary
ASPCAP: The Apogee Stellar Parameter and Chemical Abundances Pipeline
Ana E. García Pérez, Carlos Allende Prieto, Jon A. Holtzman, Matthew Shetrone, Szabolcs Mészáros, Dmitry Bizyaev, Ricardo Carrera, Katia Cunha, D. A. García-Hernández, Jennifer A. Johnson, Steven R. Majewski, David L. Nidever, Ricardo P. Schiavon, Neville Shane, Verne V. Smith, Jennifer Sobeck, Nicholas Troup, Olga Zamora, Jo Bovy, Daniel J. Eisenstein, Diane Feuillet, Peter M. Frinchaboy, Michael R. Hayden, Fred R. Hearty, Duy C. Nguyen, Robert W. O'Connell, Marc H. Pinsonneault, David H. Weinberg, John C. Wilson, Gail Zasowski
TL;DR
APOGEE’s H-band spectra contain extensive chemical information but require automated analysis because blended molecular features complicate interpretation. This paper describes ASPCAP, which fits observed spectra with synthetic libraries using multidimensional χ2 minimization, and reports its tested performance and scope. ASPCAP delivers atmospheric parameters and abundances for the predominantly giant APOGEE sample, while low metallicity, dwarf gravities, and imperfect LSF modeling remain important limitations.
Problem
Blended H-band features and molecular contaminants make automated, detailed, and accurate analysis necessary for APOGEE’s multidimensional stellar characterization.
Method
ASPCAP compares APOGEE spectra with synthetic spectral libraries using χ2 minimization in a multidimensional parameter space, combining FERRE with an IDL wrapper.
Results
Typically < 0.1 dex abundance precision is achieved in simulation tests and for DR12 M67 results, while typical abundance accuracy is expected to be ≲0.20 dex from reference-star comparisons.
Takeaways & Limitations
ASPCAP enables high-dimensional stellar-parameter and chemical-abundance analysis over a large APOGEE dataset, measuring up to 15 element abundances.
Takeaways & Limitations
Accuracy and precision decrease substantially at low metallicity or warm temperatures, and imperfect LSF modeling may contribute non-negligible errors.
Abstract
from arXiv · showhide
The Apache Point Observatory Galactic Evolution Experiment (APOGEE) has built the largest moderately high-resolution (R=22, 500) spectroscopic map of the stars across the Milky Way, and including dust-obscured areas. The APOGEE Stellar Parameter and Chemical Abundances Pipeline (ASPCAP) is the software developed for the automated analysis of these spectra. ASPCAP determines atmospheric parameters and chemical abundances from observed spectra by comparing observed spectra to libraries of theoretical spectra, using chi-2 minimization in a multidimensional parameter space. The package consists of a fortran90 code that does the actual minimization, and a wrapper IDL code for book-keeping and data handling. This paper explains in detail the ASPCAP components and functionality, and presents results from a number of tests designed to check its performance. ASPCAP provides stellar effective temperatures, surface gravities, and metallicities precise to 2%, 0.1 dex, and 0.05 dex, respectively, for most APOGEE stars, which are predominantly giants. It also provides abundances for up to 15 chemical elements with various levels of precision, typically under 0.1 dex. The final data release (DR12) of the Sloan Digital Sky Survey III contains an APOGEE database of more than 150,000 stars. ASPCAP development continues in the SDSS-IV APOGEE-2 survey.
1. INTRODUCTION
APOGEE provides high-resolution, near-infrared spectra for a large stellar sample, creating a need for automated analysis of blended H-band features. ASPCAP addresses this need and is documented here through software descriptions and performance tests.
- APOGEE obtained near-infrared spectra for over 150,000 stars between 2011 and 2014, with observations continuing through APOGEE-2.The survey maps the Galactic bulge, disk, and halo.
- R = 22,500 APOGEE H-band spectra provide information on at least 15 element abundances, typically with precision better than 0.1 dex.The spectra have S/N ≥100 per half-resolution element or approximately per pixel.
- H-band spectral analysis is challenging because many features are blended, including molecular-line contaminants, and the wavelength regime is less extensively studied than optical ranges.
- ASPCAP extracts up to 15 elemental abundances from APOGEE spectra across approximately 200 nm and about 10^4 wavelength points.Molecular features from CO, CN, and OH can affect spectroscopic parameters through molecular equilibrium and continuous opacity.
- More than 150,000 stars from the full three-year APOGEE database are publicly available in SDSS DR12.
- The paper describes ASPCAP software and model spectra, details FERRE and the IDL wrapper, and reports algorithm and software tests using survey data.The paper’s sections cover software structure, model spectra, minimization, bookkeeping, testing, and performance.
2. OVERALL ASPCAP STRUCTURE
ASPCAP combines an IDL wrapper with the FERRE optimization code to compare observed APOGEE spectra with synthetic libraries. Its iterative workflow first determines atmospheric parameters and selected broad abundances, then fits individual elemental abundances in sensitive spectral windows.
- ASPCAP has two main functional components: the fortran90 FERRE optimizer and an IDL wrapper for preprocessing, bookkeeping, and data handling.The wrapper prepares input spectra and performs multiple calls to FERRE.
- FERRE interpolates a synthetic spectral grid and identifies the atmospheric parameters and abundances yielding the best fit to each observed spectrum.
- The extracted parameters are followed by individual abundance determinations using spectral windows selected for sensitivity to each element.
- The first fitting step estimates Teff, log g, ξt, and [M/H] from the full APOGEE spectrum while allowing C, N, and α-element abundances to vary.The α-elements are O, Mg, Si, S, Ca, and Ti.
3. MODEL SPECTRA
ASPCAP models APOGEE spectra with synthetic libraries built from model atmospheres, atomic and molecular data, radiative-transfer synthesis, and instrumental broadening. The libraries are compressed and searched over stellar-parameter grids, with microturbulence reduced through an empirical gravity relation.
- Model atmospheres and synthesis: ASPCAP synthetic spectra are generated by solving radiative transfer across model-atmosphere grids for the APOGEE H-band wavelength regime.The models use APOGEE ATLAS9 atmospheres, with details supplied for atmospheres, line lists, synthesis, and databases.
- Atomic and molecular data: The atomic and molecular line lists combine Kurucz data with laboratory updates and were fine-tuned using the Sun and Arcturus.
- Spectral synthesis: Synthetic spectra are produced with ASSϵT at very high resolution and then smoothed to account for instrumental broadening.DR10 used a Gaussian kernel, whereas DR12 used a more realistic empirical kernel.
- Library compression: PCA compression divides libraries into 30 wavelength intervals, retains 30 components per interval, and reduces library size by nearly a factor of ten.The retained coefficients total 900 across the wavelength intervals.
- Search grids: DR12 searches separate PCA-compressed synthetic-spectrum grids spanning 3500–8000 K from early-M to F spectral classes.The parameter space includes Teff, log g, [M/H], and usually C, N, and α-element abundances.
- Fitting parameters: Microturbulence is treated through a linear relation with surface gravity to reduce fitting dimensionality, with separate DR10 and DR12 relations.For stars hotter than 6000 K in DR10, microturbulence was fixed at 2 km s−1.
4. FERRE
FERRE is ASPCAP’s optimization engine, matching observed spectra to synthetic libraries and minimizing χ2 across stellar parameters and abundances. It uses weighted spectral comparisons, multi-start searches, interpolation, masking, and curvature-based error estimates to derive global and elemental parameters.
- Algorithm: ferre is a fortran90 optimization code that finds the model-spectrum parameters best matching each observed spectrum.It can optimize several spectra in parallel using OpenMP and supports configurable search algorithms and interpolation schemes.
- Algorithm: Observed spectra are matched against synthetic spectra using χ2 as a merit function, with weights derived from flux uncertainties.Weights are increased in regions severely affected by sky emission and incorporate spectral-feature sensitivity for abundance measurements.
- Algorithm: Global fits use 12 searches initialized across a six-dimensional parameter space, and the lowest-converged χ2 solution is accepted.The searches use the Nelder & Mead algorithm and stop when the χ2 standard deviation at simplex test points falls below 10^-4.
- Algorithm: Cubic Bézier interpolation evaluates model fluxes between grid nodes, while wavelength-dependent weights or masks exclude bad and contaminated pixels.ASPCAP used cubic Bézier interpolation for both DR10 and DR12.
- Derivation of Elemental Abundances: Elemental abundances are obtained in additional fits that vary one abundance dimension while holding other global-fit parameters fixed over element-specific spectral windows.Pixel weights are derived from synthetic-flux derivatives across metallicities, reduced for sensitivity to other elements, and adjusted using model agreement with Arcturus.
- Derivation of Elemental Abundances: The abundance analysis uses dozens of features for C, N, O, and Fe but only a handful for several other elements among the 15 measured species.The selected windows and weights are documented in Table 3, and visual inspection removes additional problematic regions.
5. IDL WRAPPER
The IDL wrapper prepares APOGEE spectra, coordinates repeated FERRE optimizations, and manages data handling before and after fitting. It normalizes and quality-controls spectra, organizes element-specific processing, and packages calibrated outputs.
- IDL Wrapper: The IDL wrapper performs ASPCAP’s non-optimization tasks, including input preparation, bookkeeping, and multiple calls to FERRE.It is the companion component to the FERRE optimization code.
- Data Preparation: Processing is organized by APOGEE field and conducted independently for each stellar spectral class.Fields are identified by Galactic coordinates or a unique location ID.
- Data Preparation: Observed spectra are shifted to the stellar rest frame and placed on the synthesis wavelength scale using sinc interpolation.The wavelength correction uses stellar radial velocities estimated by the APOGEE data-reduction pipeline.
- Data Preparation: ASPCAP compares continuum-normalized observed and synthetic fluxes, normalizing each detector independently with repeated sigma-clipped polynomial fits.Normalization reduces effects associated with reddening and instrumental response while maintaining consistency between observations and libraries.
- Data Preparation: Bad pixels and regions affected by instrumental, atmospheric, interstellar, or sky-emission contamination are excluded from continuum fitting and χ2 evaluation.The selected chemical-abundance windows are shown in Figure 3 and listed in electronic Table 3.
- Output Processing: The wrapper writes FERRE inputs, submits jobs, organizes outputs, assigns quality flags, performs calibrations, and packages results into FITS files.Outputs include normalized spectra, best-fitting models, parameters, abundances, and covariance matrices.
6. TESTS ON SIMULATED DATA
Synthetic-spectrum tests show that ASPCAP generally recovers input parameters well, with performance depending on interpolation, PCA compression, noise, stellar properties, and spectral modeling assumptions.
- PCA compression: At typical APOGEE metallicities, stellar parameters and C, N, and α-element abundances were well recovered, and S/N = 100 noise did not compromise overall quality.Compression was more problematic at lower metallicities, particularly for carbon and nitrogen.
- Interpolation: Cubic interpolation produced the best overall results, while quadratic and cubic interpolation reduced dispersion by about two times relative to linear interpolation for most parameters.ASPCAP therefore uses cubic interpolation in its analysis.
- PCA compression: PCA compression produced dispersions of approximately 70 K for Teff, 0.15 dex for log g, 0.2 dex for [C/M], and 0.45 dex for [N/M].Metallicity, Teff median offsets, and log g median offsets remained insignificant compared with other uncertainty sources.
- Noise: Noiseless spectra yielded uncertainties of 21 K in Teff and less than 0.04 dex for other parameters, while nitrogen remained uncertain at σ([N/M]) ≈ 0.17 dex.These results estimate pure systematic uncertainties.
- Noise: At S/N = 25, recovery was acceptable except for [N/M], whose dispersion was 0.28 dex; at S/N = 50, σ was below 30 K for Teff and 0.1 dex for other parameters except nitrogen.Results at S/N = 100 or 200 were similar to the associated noiseless test, while benefits above S/N = 100 were marginal.
- Parameter dependence: Uncertainties increased at [M/H] ≲ −1.0, for warm spectra, and at dwarf gravities, with C, N, and microturbulence uncertainties sometimes exceeding 0.1 dex.Most APOGEE stars lie at [M/H] ≳ −1.0, but the spectroscopic information content decreases at low metallicity or warm temperatures.
- Error estimates: Curvature-matrix internal errors were comparable to or better than multiple-search estimates, and large input-output differences generally corresponded to large internal error estimates.At low metallicity, internal abundance and microturbulence errors could be excessive, with better agreement at higher S/N.
7. EXAMPLES WITH REAL DATA
Tests on real giant-star spectra show generally good agreement between ASPCAP and literature abundances, while surface gravity and selected elements exhibit systematic discrepancies. The comparisons also expose sensitivities to spectral resolution, stellar parameters, and atomic data.
- Test sample: High-quality FTS and 1-m H-band spectra of Arcturus, β And, δ Oph, and µ Leo were analyzed against literature reference values.The sample spans 3825–4550 K, log g 0.90–1.70, and [Fe/H] −0.47 to +0.31.
- Stellar parameters: +0.68 dex is the surface-gravity discrepancy for µ Leo, compared with approximately 0.2–0.3 dex for β And and δ Oph.ASPCAP surface gravities are systematically larger than the literature values in this sample.
- Stellar parameters: Within 0.4 km s−1, microturbulent velocities generally agree with Smith et al. for the three comparison stars despite an effective-temperature trend.Arcturus is an exception in the 1-m analysis, where the microturbulent-velocity offset is −0.43 km s−1.
- Stellar parameters: 0.04, 0.01, and 0.14 dex are the metallicity differences for β And, δ Oph, and µ Leo, respectively, while Fe abundances agree even better.Agreement in stellar parameters remains good at the higher metallicities represented by these stars.
- Abundances: ≳0.15 dex discrepancies recur for N, O, K, and Si in some stars, but only Si systematically exceeds the Smith et al. values beyond expected errors.The repeated Si offset supports a contribution from differences in adopted atomic data; K offsets vary between stars.
- Abundances: < 0.10 dex dispersion characterizes abundance differences from Smith et al. (2013), with especially good agreement for Ca, Fe, and Ni except µ Leo.The overall abundance comparison is described as good, although Si is the most discrepant element.
8. CONCLUSIONS
ASPCAP derives stellar parameters and chemical abundances by matching APOGEE spectra to synthetic templates through multidimensional χ2 minimization. The paper documents its components, reports successful tests on simulations and observations, and identifies important performance limitations and ongoing development areas.
- Pipeline functionality: ASPCAP derives stellar parameters from the full spectrum before determining individual abundances in element-optimized spectral windows.The analysis uses χ2 minimization in a multidimensional parameter space.
- Pipeline architecture: ASPCAP combines model spectral libraries, the ferre best-fit search, and an IDL wrapper for bookkeeping and data pre- and post-processing.The paper presents configurations used in DR10 and DR12, whose algorithms worked well with simulations and observations.
- Performance: Typically < 0.1 dex random abundance uncertainties are expected from simulations and the DR12 M67 results, while accuracy is typically ≲0.20 dex from reference-star comparisons.These are separate random-uncertainty and accuracy expectations.
- Limitations: Detailed line-spread-function modeling is important because poor LSF matching can produce appreciable systematic effects.DR12 used an empirical LSF, whereas DR10 used a Gaussian LSF of constant resolving power.
- Limitations: ASPCAP performance is poorer for metal-poor or warm spectra because these regions lack information, motivating models with fewer parameters.PCA compression may also affect low-metallicity spectra with [M/H] < −1 and requires further investigation.
- Limitations: Nitrogen abundances have significant uncertainties because CN lines are only modestly sensitive to changes in nitrogen abundance.This limits the reliability of inferred nitrogen abundances relative to better-constrained elements.
- Future development: ASPCAP development continues through expanded parameter coverage, abundance upper limits for undetected lines, improved LSF modeling, and possible independent elemental-abundance fits.These efforts continue in the context of APOGEE-2.