A Low-Dimensional Chemical Representation of High-Resolution Stellar Disk Spectra: Residual Abundance Signatures

Abstract

Stellar abundances are generated by complex chemical enrichment pathways yet show strong correlation structure. We test how well high-quality spectra of low-α disk stars can be reconstructed using low-dimensional models. By validating model performance in flux space, we quantify and locate where residual information persists. Using HARPS spectra (R = 115, 000, SNR >100) and labels derived with korg, we train generative models (the cannon) conditioned on stellar parameters Teff, log (g), [M/H], vsin i, and vmic, and three abundance labels, [Ti/H], [Y/H], and [Nd/H], representing distinct nucleosynthetic families. Evaluating model performance on unseen test stars through flux residuals, the model reproduces the global spectra to a median residual amplitude of 0.3%. Deviations in 15 absorption features for unlabelled elements are typically ≤0.7%. Residuals of ∼1% correspond to abundance differences of 0.01–0.05 dex. Degrading the spectra to medium resolution (R = 28, 000) decreases the median residual amplitude from 0.3% to 0.2%, redistributing the residual signal over the instrumental line profile. Our publicly available model reproduces features from 29 elements to sub-percent precision across −1.13 < [M/H] < 0.29. This demonstrates that the dominant chemical variation encoded in low-α disk spectra can be captured by a low-dimensional model, while weaker chemical information remains encoded in flux residuals. High-resolution, high-SNR spectra (R ∼ 100, 000, SNR >100) are essential to resolve the hierarchical layers of chemical information beyond this low-dimensional structure. The model provides a benchmark for assessing the chemical information accessible to current surveys like 4MOST and GALAH 2, as well as upcoming facilities like VLT/HRMOS and GMT/G-CLEF.