Skip to main content
Back to timeline
arXivSource publication:

COMPASS folds yield-table comparison and parameter inference into one diffusion model, and Milky Way data favor a NuGrid plus TNG hybrid

Synopsis

COMPASS puts two jobs of Galactic chemical evolution into one conditional score-based diffusion framework: comparing competing stellar nucleosynthesis yield tables and inferring galactic parameters, learning both posterior and likelihood directly from forward simulations. In controlled tests it identifies the generating yield table and recovers ground-truth parameters with accuracy comparable to existing simulation-based inference, at orders-of-magnitude lower cost than traditional sampling. Applied to high-precision abundances of 69 solar-type stars, it favors NuGrid asymptotic giant branch yields combined with the core-collapse supernova yields used in IllustrisTNG, and infers a high-mass initial mass function slope steeper than the canonical Salpeter value.

Source-provided article image: A COMPASS to the Galaxy: Joint nucleosynthetic yield set comparison and parameter inference in Galactic chemical evolution
arXiv

Interpretation

Galactic chemical evolution models have typically fixed one nucleosynthesis yield table and then inferred parameters, leaving yield-table uncertainty outside the inference. Earlier simulation-based inference work estimated parameters within a fixed physical model and did not unify parameter inference with Bayesian comparison among competing yield tables. On controlled mock data generated with three core-collapse supernova yield tables, COMPASS assigns the highest model weight to the true generating model for the vast majority of single-star observations, and the combined weight of the correct model approaches unity after only a few observations.

The same conditional score-based diffusion model supplies both a parameter posterior and a likelihood, so candidate yield tables can be ranked by maximized likelihood. The diffusion model uses a transformer to parameterize the gradient of the log-density of the data distribution, the score, and switches between neural posterior estimation and neural likelihood estimation through a condition mask. On C, O, Mg, Si and Fe abundances of 69 HARPS solar-type stars from Nissen et al. 2020, the combination of NuGrid asymptotic giant branch yields with the TNG core-collapse supernova yields reaches a cumulative posterior approaching unity, and an independent neural classifier that does not use the COMPASS likelihood also prefers the same combination by a large margin.

Under the preferred yield table, the inferred high-mass initial mass function slope is steeper than the canonical Salpeter value, while the Type Ia supernova normalization scatters around the prior mean across yield tables. The posteriors of the six highest-ranked yield combinations are mutually inconsistent at several standard deviations, making the yield-table choice the dominant source of uncertainty on these parameters. The preferred combination's IMF slope and SNIa normalization are tabulated in the results table, as are the IMF-slope span and statistical uncertainties across the six combinations; those specific numbers are stripped from the readable text here, so only the directional conclusion is kept.

Perspective

The demonstrated setting is the one-zone Chempy chemical evolution model applied to C, O, Mg, Si and Fe abundances of 69 solar-type stars from Nissen et al. 2020, with stellar ages as conditioning variables and missing elements handled by attention masking; the controlled validation uses mock data generated with three core-collapse supernova yield tables.

The model weights are Akaike weights at the maximum a posteriori parameters, not evidence ratios, and omit the prior volume of each model; single-star posteriors are combined under a Gaussian approximation, so shared systematic errors grow in significance with the number of stars; COMPASS posteriors are mildly under-confident for the global parameters; under yield-table misspecification or with IllustrisTNG hydrodynamical data, all tested methods are biased by several standard deviations and credible regions are not reliable in that regime.