Towards model-free stellar chemical abundances. Potential applications in the search for chemically peculiar stars in large spectroscopic surveys

TL;DR

This work tackles the challenge of extracting stellar chemical abundances from spectra without relying on extensive labeled catalogs or imperfect atmosphere models. It introduces a self-supervised, disentangled representation learning framework based on a variational autoencoder with element-specific decoders, enforcing that latent features correspond to $[Fe/H]$, $[C/Fe]$, and $[\alpha/Fe]$ while disentangling non-chemical factors like $T_{\rm eff}$ and $\log g$. The model demonstrates high-quality reconstruction on synthetic low-resolution spectra and shows that latent axes strongly correlate with the intended abundances ($r$ values of $0.92$, $0.92$, and $0.82$), enabling robust flagging of chemically enhanced or depleted stars (e.g., $\alpha$PMP and $\text{CEMP}$) with high precision. It also introduces selective gradient flow to prevent cross-talk between latent factors, and relies on a Gaussian prior to facilitate outlier detection in a principled way. The approach holds promise for scalable chemical tagging in large spectroscopic surveys and offers a complementary path to model-based abundance inference, with demonstrated potential on real data and clear avenues for extension to additional elements and higher-resolution datasets.

Abstract

Chemical abundance determinations from stellar spectra are challenged by observational noise, limitations in stellar models, and departures from simplifying assumptions. While traditional and supervised machine learning methods have made remarkable progress in estimating atmospheric parameters and chemical compositions within existing physical models, these factors still constrain our ability to fully exploit the vast data sets provided by modern spectroscopic surveys. We aim to develop a self-supervised, disentangled representation learning framework that extracts chemically meaningful features directly from spectra, without relying on externally imposed label catalogs. We build a variational autoencoder-based representation learning model with physics-inspired structure: multiple decoders each focus on spectral regions dominated by a particular element, enforcing that each latent dimension maps to a single abundance. To evaluate the potential application of our framework, we trained and validated the model on low-resolution, low signal-to-noise synthetic spectra focusing on $\rm [Fe/H]$, $\rm [C/Fe]$, and $\rm [α/Fe]$. We then demonstrate how the trained model can be used to flag stars as chemically enhanced or depleted in these abundances based on their position within the latent distribution. Our model successfully learns a representation of spectra whose axes correlate tightly with the target abundances ($r=0.92\pm0.01$ for $\rm [Fe/H]$, $r=0.92\pm0.01$ for $\rm [C/Fe]$, $r=0.82\pm0.02$ for $\rm [α/Fe]$). The disentangled representations provide a robust means to distinguish stars based on their chemical properties, offering an efficient and scalable solution for large spectroscopic surveys.

Figures (11)

• Figure 1: Main model. A spectrum is given as input. An encoder maps it to the latent space, divided into chemical latent space $z=(z_{\rm M},z_{\rm C},z_{\rm \alpha})$, and auxiliary latent space $\vec{w}$. Three decoders map different components of $\vec{z}$, along with $\vec{w}$, to separate regions of the input spectrum. A linear transform is applied to the auxiliary space to obtain ($T_{\rm eff},\log g$). The discriminator is trained to predict the non-chemical parameters ($T_{\rm eff},logg$) from $\vec{z}$, while the encoder is trained to make this impossible for the discriminator to predict these parameters from $\vec{z}$ alone. Grey dotted arrows represent the absence of gradient propagation (see Sect. \ref{['sec:gradients']}). At inference time the decoders, the discriminator and the linear transform module are discarded.
• Figure 2: Probability distribution functions (PDFs) for the chemical abundances used to sample the stellar properties in our dataset,with lighter colors indicating higher probability densities. Left panel: PDF of ${\rm [C}/{\rm Fe]}$ versus ${\rm [Fe}/{\rm H]}$. Right panel: PDF of ${\rm [\alpha}/{\rm Fe]}$ versus ${\rm [Fe}/{\rm H]}$. In both panels, the pink contours show the abundance distributions of halo stars from the APOGEE survey, as provided by the astroNN catalog. APOGEE does not reach the lowest metallicities present in our simulated dataset, which explains the differences at low ${\rm [Fe}/{\rm H]}$.
• Figure 3: Comparison of original (before noise perturbation, blue) and reconstructed spectra (pink) for three chemical types: an $\alpha$-poor, metal-poor star (top row), a carbon-rich, metal-poor star (middle row), and a solar-like star (bottom row). The residuals (reconstructed minus original) are shown in red in the bottom subpanels. Shaded regions highlight spectral domains handled by the different decoders, as indicated in the legend.
• Figure 4: Contour plot of latent features (from left to right, $z_{\mathrm M}, z_{\mathrm \alpha}, z_{\mathrm C}$) and their corresponding chemical abundances. The scatter points represent individual data points, and the contour lines represent data density, with lighter contours indicating regions of higher density. Straight lines show linear fits to the latent-abundance relations for stars in three $T_{\rm eff}$ bins, as indicated in the legend.
• Figure 5: Distribution of stars in the true chemical space. Left panel: $[\mathrm{C}/\mathrm{Fe}]$ vs $[\mathrm{Fe}/\mathrm{H}]$. Right panel: $[\alpha/\mathrm{Fe}]$ vs $[\mathrm{Fe}/\mathrm{H}]$. The color represents the predicted class, with blue indicating stars not flagged as enhanced/depleted, and pink indicating those flagged as enhanced/depleted. The pink-highlighted regions indicate the approximate boundaries for $\alpha$PMP and CEMP stars, as described in Sect. \ref{['sec:ad']}, where $\alpha$PMP stars are defined by $[\alpha/{\rm Fe}]<0.0$ and $[{\rm Fe}/{\rm H}]<-1.0$, and CEMP stars by $[{\rm C}/{\rm Fe}]>0.7$ and $[{\rm Fe}/{\rm H}]<-1.0$.