TwinSpecNet: Extending APOGEE’s chemical reach to low-S/N spectra via empirical paired learning
Weijia Sun ID, Cristina Chiappini ID, Samir Nepal ID · 2026
Read TwinSpecNet: Extending APOGEE’s chemical reach to low-S/N spectra via empirical paired learning by Weijia Sun ID, Cristina Chiappini ID, Samir Nepal ID (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-26491.txt.
Introduction
Sun, Chiappini, and Nepal built TwinSpecNet, a paired-learning framework that rescues chemical labels from low signal-to-noise APOGEE spectra. Faint distant populations in the bulge, halo, and satellites still carry recoverable chemistry, but standard pipelines grow imprecise or skip them. Training on low- and high-quality spectral twins of the same stars, the Vision Transformer rebuilds clean flux and predicts parameters with calibrated uncertainty. Label scatter fell for faint visits, with residual scatter of about 0.06 dex in surface gravity, and age precision improved for APOKASC giants.
Catalog record
| CREATOR | Weijia Sun ID, Cristina Chiappini ID, Samir Nepal ID |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74766 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.26491
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
TwinSpecNet: Extending APOGEE’s chemical reach to low-S/N spectra via empirical paired learning Weijia Sun ID, Cristina Chiappini ID, Samir Nepal ID Abstract Context. Large spectroscopic surveys rely on automated pipelines to deliver homogeneous stellar labels, but a substantial fraction of observations are at low signal-to-noise ratio (S/N), where label estimates become imprecise or are omitted. In APOGEE, these low-S/N spectra visits sample faint and distant populations—the bulge, outer halo, and satellite systems—yet still encode recoverable chemical information. Aims. We present TwinSpecNet (TSN), a paired-learning framework that exploits APOGEE’s multi-visit observing strategy: by training on empirical low-/high-S/N spectral twins of the same stars, TSN learns to suppress stochastic noise while preserving the ASPCAP label scale. Methods. TSN employs a Vision Transformer encoder with dual objectives: reconstructing high-S/N flux from low-S/N visits and predicting stellar parameters and abundances with calibrated uncertainties. Results. TSN reduces label scatter relative to visit-level ASPCAP for S / N < 60 \mathrm{S/N}<60 visits. TSN reproduces the ASPCAP scale with residual scatters of σ ≃ 19 K \sigma\simeq 19\,\mathrm{K} in T eff T_{\mathrm{eff}} , σ ≃ 0.06 dex \sigma\simeq 0.06\,\mathrm{dex} in log g \log g , and σ ≃ 0.03 dex \sigma\simeq 0.03\,\mathrm{dex} in [ Fe / H ] \rm[Fe/H] . TSN tightens intra-cluster abundance dispersions, recovers cleaner chemical sequences in inner-disk and bulge and satellite samples, and improves C/N-based age precision for APOKASC giants from 1.70 to 1.49 Gyr. Conclusions. By learning survey-specific noise patterns from repeated observations, TSN demonstrates how empirical paired learning can extend the chemical reach of existing spectroscopic data, providing a template applicable to other multi-visit surveys. 1 Introduction High-resolution stellar spectroscopy underpins Galactic archaeology by enabling large-scale inference of stellar parameters and detailed chemical abundances ( Feltzing & Chiba 2013 ; Jofré et al. 2019 , e.g.) . In combination with astrometry and kinematics, such chemo-dynamical data provide constraints on the formation and evolution of the Milky Way disc, bulge, and halo ( Queiroz et al. 2021 ; Binney & Vasiliev 2024 , e.g.) . To enable population-level studies, modern multiplexed surveys invest heavily in automated pipelines that deliver homogeneous stellar labels ( Majewski et al. 2017 ; De Silva et al. 2015 , e.g.) . In APOGEE, the APOGEE Stellar Parameters and Chemical Abundances Pipeline (ASPCAP) infers atmospheric parameters and abundances by fitting synthetic spectral grids ( García Pérez et al. 2016 ) . However, pipeline performance degrades sharply at low signal-to-noise ratios (S/N), where spectral features become difficult to distinguish from instrumental noise and reduction residuals. Key Galactic components—the distant bulge and inner disc behind high extinction, the outer halo, and satellite systems—are probed primarily by faint targets that yield low S/N spectra. Recovering chemically informative labels from these observations is therefore critical for studies of bulge and halo formation and for identifying accreted substructure in the Milky Way’s outskirts ( Yuan et al. 2020 ; Horta et al. 2023 , e.g.) . Yet APOGEE has accumulated millions of individual visit spectra at low S/N, where single-visit observations are noise dominated and the corresponding co-added “combined” spectra, although improved, do not always reach the high-S/N regime preferred for precise abundance work. In this regime, grid-fitting pipelines return imprecise stellar labels with large uncertainties or unreliable flags, and visit spectra are often excluded from abundance-driven analyses. Nevertheless, recent studies have demonstrated that low-S/N spectra can in fact yield substantial new insights when analysed with advanced machine-learning techniques. In particular, the application of convolutional neural networks to Gaia RVS spectra has revealed that even at low S/N, accurate stellar parameters can be recovered, significantly enhancing the scientific return of the RVS dataset ( Guiglion et al. 2024 ) . These advances open the door to the exploitation of large, precise statistical samples, enabling population-level studies that were previously inaccessible, as illustrated by the new results obtained in Nepal et al. 2024 and Nepal et al. 2025 . APOGEE is a high-resolution near-infrared spectroscopic survey designed to map the chemo-dynamical structure of the Milky Way using H-band stellar spectra ( Zasowski et al. 2013 ; Majewski et al. 2017 ) . APOGEE provides a unique empirical advantage through its co-added-visit observing strategy. Many stars have been observed multiple times, and for a subset the combined spectrum reaches high S/N. These repeated observations define empirical low-/high-S/N pairs for the same star observed with the same instrument, which can be used to supervise denoising without assuming an explicit noise model. This paired-learning paradigm is closely related to Noise2Noise-style training in imaging ( Lehtinen et al. 2018 ) , but here the “twin” is an actual higher-quality spectrum of the same astrophysical source. Such empirical spectral twins encode which spectral structures are stable (astrophysical signal) versus those that vary stochastically (instrumental noise and reduction residuals), providing direct supervision for noise suppression that is anchored to survey-specific systematics. Despite this opportunity, existing approaches to low-S/N label inference have not fully exploited APOGEE’s paired structure. Data-driven label-transfer methods (e.g. The Cannon ; Ness et al. 2015 ) and deep-learning estimators (e.g. StarNet; Fabbro et al. 2018 ) can infer labels from noisy spectra and have been shown to perform well down to relatively low S/N. However, purely discriminative spectrum-to-label models do not provide an explicit denoised spectrum for inspection, and can regress toward the training-set mean when the inputs are dominated by measurement noise ( Ting 2025 ) . Additionally, APOGEE visit spectra can exhibit persistence, detector artefacts, and reduction residuals that vary from visit to visit ( Holtzman et al. 2015 ; Jönsson et al. 2020 ) , introducing correlated structure beyond simple per-pixel Poisson noise that discriminative models may not capture effectively. A complementary strategy that addresses these limitations is to couple label inference to an empirical spectral reconstruction objective. Physically motivated forward models such as The Payne ( Ting et al. 2019 ) predict spectra as a function of labels, enabling residual inspection alongside label inference. A data-driven variant, the dd-Payne ( Xiang et al. 2019 ) , replaces the synthetic spectral grid with an emulator trained on observed reference spectra, thereby avoiding biases associated with incomplete or imperfect physical models. More recently, hybrid architectures couple label prediction to auxiliary spectral reconstruction ( Li et al. 2024 , e.g. cross-resolution reconstruction,) . However, these approaches typically rely on synthetic models or cross-dataset mappings that may introduce physical priors not present in the data. For low-S/N stellar spectra, a practical requirement is therefore to suppress stochastic noise while remaining anchored to an empirically validated reference spectrum from the same survey and a well-defined label scale. By training on empirical visit–combined spectral twins, such a model can learn survey-specific noise patterns and reduction residuals while preserving APOGEE’s internal reference frame. In this work we introduce TwinSpecNet (TSN), an empirical paired-learning framework for rescuing low-S/N APOGEE visit spectra. TSN uses a Vision Transformer (ViT) encoder trained sequentially with two objectives: (i) a paired denoising objective that maps low-S/N visits to high-S/N combined spectra, and (ii) a probabilistic label model that predicts stellar parameters and abundances with calibrated uncertainties on the ASPCAP scale. We assess whether paired learning can extend the effective chemical reach of APOGEE. Specifically, we test whether TSN (a) improves label precision