Code acompaning the manuscript "Toward model-free stellar chemical abundances", aa55376-25
This repository contains the implementation of the physics-constrained, self-supervised representation learning framework introduced in the paper. The method combines variational autoencoders with physically motivated inductive biases to learn interpretable, element-specific latent dimensions directly from stellar spectra, enabling:
Disentangled chemical representations
Noise-robust latent features aligned with abundance variations
Physically meaningful anomaly detection
A step toward model-free chemical abundance inference
The paper introduces and validates the framework on synthetic spectra and demonstrates applications to identifying chemically anomalous stars (e.g., α-poor, metal-poor, and CEMP stars).
If you use this code, please cite the paper:
“Toward model-free stellar chemical abundances” A&A, accepted (aa55376-25). Preprint: https://arxiv.org/abs/2511.09733