A Machine-Learning Compositional Study of Exoplanetary Material Accreted Onto Five Helium-Atmosphere White Dwarfs with $\texttt{cecilia}$

TL;DR

The paper demonstrates a scalable, Bayesian, ML-driven approach to infer white dwarf atmospheric parameters and accreted material compositions from optical spectra. By applying the cecilia pipeline to five He-atmosphere polluted WDs with SDSS and Keck/ESI data, the authors recover 2–6 detected metals per star with ~0.20 dex precision and infer pollutant compositions largely consistent with CI chondrites, including significant oxygen excess in two cases. The work validates the ML method against classical spectroscopy and discusses limitations due to limited metal detections and the lack of UV data, while outlining concrete pathways to extend cecilia for population-wide studies of exoplanetary debris in the era of large spectroscopic surveys.

Abstract

We present the first application of the Machine Learning (ML) pipeline $\texttt{cecilia}$ to determine the physical parameters and photospheric composition of five metal-polluted He-atmosphere white dwarfs without well-characterised elemental abundances. To achieve this, we perform a joint and iterative Bayesian fit to their $\textit{SDSS}$ (R=2,000) and $\textit{Keck/ESI}$ (R=4,500) optical spectra, covering the wavelength range from about 3,800Å to 9,000Å. Our analysis measures the abundances of at least two $-$and up to six$-$ chemical elements in their atmospheres with a predictive accuracy similar to that of conventional WD analysis techniques ($\approx$0.20 dex). The white dwarfs with the largest number of detected heavy elements are SDSS J0859$+$5732 and SDSS J2311$-$0041, which simultaneously exhibit O, Mg, Si, Ca, and Fe in their $\textit{Keck/ESI}$ spectra. For all systems, we find that the bulk composition of their pollutants is largely consistent with those of primitive CI chondrites to within 1-2$σ$. We also find evidence of statistically significant ($>2σ$) oxygen excesses for SDSS J0859$+$5732 and SDSS J2311$-$0041, which could point to the accretion of oxygen-rich exoplanetary material. In the future, as wide-field astronomical surveys deliver millions of public WD spectra to the scientific community, $\texttt{cecilia}$ aspires to unlock population-wide studies of polluted WDs, therefore helping to improve our statistical knowledge of extrasolar compositions.

Figures (14)

• Figure 1: Median-normalised optical spectra of the five polluted WDs in our sample (left: SDSS with an assumed fixed resolving power of R$=$2,000 (in vacuum); right: Keck/ESI with an assumed fixed R$=$4,500 in air). The stellar fluxes and their corresponding uncertainties are shown, respectively, in blue and red, while cecilia's best-fit models are presented in dark green. The data gaps correspond to the discarded wavelength regions described in Section \ref{['sec:paper3_data_treatment']}, including the strongest He I lines. The grey dashed vertical lines denote the 200 Å spectral windows used during cecilia's training and optimisation routine.
• Figure 2: A summary of cecilia's methodology for estimating the main astrophysical properties (or labels) of polluted He-rich WDs from multiple spectroscopic observations. We refer the reader to BadenasAgusti:2024 for a more comprehensive description of the pipeline.
• Figure 3: A selection of spectral windows showing cecilia's best-fit RV-shifted MCMC model (in green) for the median-normalised Keck/ESI spectrum of SDSS J0859$+$5732 (in light blue; in air wavelengths). For reference, we also include cecilia's predictions when modifying the abundances of the detected elements by $\pm1$$\sigma_{\rm{tot}}$ (red and orange). The green labels show all the detected elements, defined as those with $\sigma_{\rm{stat, MCMC}}$$\leq$$\sigma_{\rm detection}=0.10$ dex and at least one visible absorption line in the spectrum. We note that cecilia struggles to model the depth of the Mg line at about 4,481 Å (see panel d). This is not the case for other Mg lines, which are fitted reasonably well by our code (e.g. panels g and j). Such behaviour may be an example of underestimated errors due to the high level of red (i.e. correlated) noise in panel d. In Section \ref{['sec:paper3_future_work']}, we discuss how cecilia can be improved to address this problem.
• Figure 4: MCMC corner plot for SDSS J0859$+$5732. The off-diagonal plots illustrate the two-dimensional marginalised posterior distributions of the free model parameters, while the histogram panels along the diagonal show their one-dimensional marginalised distributions together with their median value and $1\sigma$ confidence interval. This Figure does not show our results for cecilia's undetected elements. It also excludes the RV shifts because cecilia does not correct for the barycentric motion and gravitational redshift of the white dwarf.
• Figure 5: Compositional properties of the WD pollutants during build-up, steady-state, and decaying phase (first, second, and third columns, respectively). The top panels show the Mg-normalised linear abundance ratios of the accreted material relative to those of CI chondrites (red line; Alexander:2019a_noncarbonaceousAlexander:2019b_carbonaceous). The bottom panels present the percent metal mass fractions for systems with simultaneous detections of at least the four major rock-forming elements (O, Fe, Si, Mg), with Ca values scaled by a factor of 10 for clarity. In each plot, we only include cecilia's detected elements (excluding Mg, our reference metal), with error bars reflecting a total assumed abundance error of $\sigma_{\rm{tot}}$=$\sigma_{\rm floor}=0.20$ dex (see Table \ref{['tab:paper3_mcmc_results']}). For comparison, we also show the properties of bulk Earth (black '$\oplus$' marker; Allegre:2001), comet Halley (black '$\Diamond$' marker; Jessberger:1988), and CI chondrites (black '$\bf{+}$' marker, bottom panels only). Heavy elements are sorted by increasing condensation temperature $T_{\rm cond}$Lodders:2003.