Abstract
We demonstrate a physics-blind, data-driven method for discovering nucleosynthetic archetypes from heterogeneous stellar elemental abundance measurements. By applying non-negative matrix factorization (NMF) and low-rank matrix completion to the Hypatia Catalog and JINAbase, we extract "endmembers"—archetypal elemental abundance patterns—and impute missing measurements. Our results indicate that this analysis, relying solely on the incomplete chemical mixing inherent in star formation, can discern nucleosynthetic archetypes associated with asymptotic giant branch (AGB) stars, thermonuclear supernovae, and core-collapse supernovae with r-process signatures.
Introduction
The chemical enrichment of the universe is a complex process driven by various stellar sources. Traditionally, astronomers have relied on "forward modeling," comparing complex physics simulations to observational data. However, the overlapping signatures of different processes and the sparsity of observational catalogs—where many stars lack measurements for specific elements—complicate this approach.
This study proposes a "backward modeling" alternative. By treating the observed chemical composition of stars as a mixture of fundamental "archetypes," we can use statistical techniques to recover the underlying sources of elements purely from the data, without prior astrophysical assumptions.
Methods
The methodology centers on Non-negative Matrix Factorization (NMF) combined with low-rank matrix completion. NMF is ideal for chemical abundances because it ensures all results are non-negative, mirroring the physical reality that a source cannot contribute a "negative" amount of an element.
To handle the "sparse" nature of astronomical data, we employ matrix completion to fill in missing values based on the global correlations within the dataset. The optimization process uses Nesterov’s accelerated gradient descent, and the robustness of the discovered patterns is validated through bootstrap procedures.
Datasets
The analysis utilized two major stellar abundance compilations:
- The Hypatia Catalog: Contains high-resolution data for ~6,000 stars in the solar neighborhood.
- JINAbase: Focuses on ~1,500 metal-poor stars, which act as chemical "fossils" from the early universe.
Results and Discussion
The core finding is that the incomplete mixing of the interstellar medium is a "feature" that allows for the recovery of distinct nucleosynthetic signatures. When the algorithm is allowed to find three archetypes (Rank $R=3$), it successfully isolates patterns corresponding to AGB stars (s-process), Type Ia supernovae (iron-peak elements), and core-collapse supernovae (alpha-elements and r-process).
In the JINAbase dataset, the inclusion of heavier elements allows the algorithm to refine the r-process signature. This demonstrates the method's capability to handle diverse datasets and extract physically meaningful patterns without being "taught" the underlying physics.
Conclusion
This "physics-blind" approach provides a powerful tool for exploring the chemical history of the Galaxy. While limited by selection biases in current catalogs, the method is expected to become increasingly effective as large-scale, homogeneous surveys like APOGEE and Gaia-ESO provide more complete datasets. This data-driven framework may eventually reveal rare nucleosynthetic pathways that challenge or extend existing theoretical models.
Summary updated