Toward model-free stellar chemical abundances

Context. 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 datasets provided by modern spectroscopic surveys. Aims. 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. Methods. We built a variational autoencoder-based representation learning model with a physics-inspired structure comprising multiple decoders, each of which focuses 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 [Fe/H],[C/Fe], and [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. Results. Our model successfully learns a representation of spectra whose axes correlate tightly with the target abundances (r= 0.92 ± 0.01 for [Fe/H], r=0.92 ± 0.01 for [C/Fe], and r=0.82 ± 0.02 for [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.