Impact of Simultaneous Stellar Modeling Uncertainties on the Tip of the Red Giant Branch for Axion-Election Coupling

TL;DR

This study presents a framework to incorporate covariances from stellar input physics into constraints on the axion–electron coupling α_{26} derived from TRGB $M_I$. By training ML emulators on grids of stellar models that vary $M$, $Y$, $Z$, and $oldsymbol{α_{26}}$ (with two fixed physics benchmarks) and employing MCMC to vary parameters simultaneously, the authors reveal that degeneracies can significantly weaken previous bounds, yielding α_{26} ≤ 0.75 and α_{26} ≤ 1.58 at 95% C.L. Their analysis shows strong covariances between metallicity, mixing length, and axion cooling, underscoring the need to reevaluate TRGB-based limits with simultaneous parameter variation. The work also notes that using bolometric luminosities, rather than $M_I$, could mitigate some theoretical uncertainties, offering a more robust path for constraining new physics with stellar observations. Overall, the methodology enables efficient exploration of high-dimensional parameter spaces and can be extended to other stellar probes of beyond-Standard-Model physics.

Abstract

We present a novel method for incorporating the effects of stellar modeling uncertainties into constraints on the axion-electron coupling constant found using the observed calibration of the tip of the red giant branch (TRGB) I band magnitude $M_I$.~We simulate grids of models with varying initial stellar mass, helium abundance, metallicity, and axion-electron coupling $α_{26}= 10^{26} g^2_{ae}/4π$ but different (fixed) mixing lengths and mass loss efficiencies.~We then train separate machine learning emulators to predict $M_I$ as a function of the varying parameters for each grid.~Our emulators enable the use of Markov Chain Monte Carlo simulations where $α_{26}$ is varied simultaneously with the stellar parameters.~One of our grids yields a bound $α_{26}\leq 0.75$ at the 95\% confidence limit, a factor of $\sim3.7$ weaker than previous bounds;~while the other grid yields $α_{26}\leq1.58$ at the 95\% confidence limit, a factor $\sim7.8$ weaker than previous bounds.~We demonstrate that the different values we find are due to covariances between stellar and axion physics that are not accounted for by single parameter variations.~Our results suggest that the bound on $α_{26}$ derived using empirical calibrations of the TRGB I band magnitude need to be reevaluated using simultaneous parameter variation.~Alternative methods that use the bolometric luminosity instead of $M_I$ are more robust because they are not reliant upon theoretical predictions of the effective temperature.

Figures (20)

• Figure 1: Plots showing the difference in the stellar evolution tracks in the HR diagram close to the TRGB (left column) and the $M_I$ values at the TRGB (right column) for different sets of varying parameters (rows). Further details are given in the text.
• Figure 2: Variation in the TRGB $M_I$ for Benchmark 1 with varying parameters indicated in the subfigures. The mean $M_I$ in each bin was found by averaging over the other parameters. The shaded error bands show the 1$\sigma$ deviation from the mean in each bin. The dark band represents the calibration of $M_I$ used by 2020ApJ...891...57F. The jaggedness is due to the sparser grid spacing outside the nominal prior range (see the main text). We remind the reader that these are the models used to train the ML emulator; physical priors are imposed in the the MCMC analysis that restrict the parameter ranges. The wider ranges are necessary for efficient training.
• Figure 3: Same as Figure \ref{['fig:degeneracy']} but for Benchmark 2.
• Figure 4: Error distributions for $M_I$, $\Delta M_I$, $(V-I)$, and $\Delta (V-I)$ from the ML regression for Benchmark 1. These distributions were calculated by subtracting the value of the ML prediction for each quantity from the value found by applying the WL code directly to the MESA outputs for each point in our grid.
• Figure 5: The same as \ref{['fig:mError_Standard']} but for Benchmark 2.