The Stellar Winds Atlas I: Current uncertainties in mass-loss rates

This paper presents a systematic comparative analysis of stellar wind prescriptions used in modern stellar evolution codes for massive stars, examining the three primary mass-loss regimes: optically thin line-driven winds, optically thick Wolf-Rayet (WR) winds, and cool supergiant winds.

Motivation and Scope: The authors note that stellar winds are a major source of uncertainty in understanding the life and deaths of massive stars and that prescriptions for stellar winds differ substantially in both their physical assumptions and implementation, making them a dominant contributor to model-to-model variation. The paper aims to provide a unified framework for interpreting theoretical discrepancies and identifying key bottlenecks in stellar-wind modeling.

Methods: The authors compare commonly-used stellar wind recipes, evaluated independently of the complexities inherent to full stellar evolution simulations, restricting analysis to closed-form, analytical mass-loss prescriptions parameterized primarily by global stellar surface quantities (L, M, Teff, Z). They compute dedicated MESA stellar models (version 24.08.1) with various wind configurations including Standard Dutch wind scheme, Strong Winds Eta (SWE), Weak Winds Eta (WWE), Slavic winds, and Strong Winds Gamma (SWG). They also introduce Argus: an automated, Large Language Model (LLM)-driven builder for the Stellar Winds Atlas that encodes the structural logic of our MESA inlist and run star extras files.

Key Findings on Optically Thin Winds: The authors find that the predicted mass-loss rates can differ by more than an order of magnitude depending on the chosen prescription. They note that "the Vink et al. (2001) and Nieuwenhuijzen & de Jager (1990) recipes yield the strongest stellar winds, with the bistability jump producing an increase of more than an order of magnitude. They also state that recent observations suggest that the Eddington parameter Γe, rather than Teff, is the primary driver of enhanced mass loss and that population synthesis codes adopting Nieuwenhuijzen & de Jager (1990), Vink et al. (2001), or Vink & Sander (2021) likely overestimate mass-loss rates in the optically thin winds regime."

Key Findings on Optically Thick Winds: The authors identify critical flaws in the criteria used to trigger the formation of a Wolf-Rayet star. They show that the commonly used Eddington factor Γe may act as an inconsistent proxy for the initiation of optically thick winds, if applied outside its domain of validity. They demonstrate that "the alternative multi-scattering eta criterion for wind efficiency is highly model-dependent, and that its current implementation in evolutionary codes sacrifices consistency with a star's instantaneous surface properties to artificially anchor the model to the input requirements of the Vink et al. (2011) parameterization. They provide recommendations for improving the multi-scattering criterion and introduce a new, model-agnostic framework for this transition that calculates wind efficiency and escape velocities based on the star's current physical state rather than hypothetical chemically homogeneous conditions."

The Cool Wolf-Rayet Problem: The authors introduce this concept as a regime where the standard physical driver of mass loss loses its consistency. They explain that the derivation of Γe is fundamentally based on the assumption that the stellar opacity is constant and dominated by a single process: Thomson scattering by free electrons in a fully ionized plasma and that this condition is only met in the hottest O-type and very early B-type stars, which have effective temperatures Teff ≳ 30 kK. Below approximately 15 kK, the recombination of hydrogen introduces an 'opacity cliff' and any model that uses the classical Γe for stars cooler than 30 kK is not correctly estimating the true radiative force acting on the stellar material. They note that no dedicated mass-loss recipe exists for cool WR stars and that current evolutionary codes must employ ad-hoc workarounds.

Key Findings on Cool Supergiant Winds: The authors find that the landscape of available prescriptions is equally divergent and that "recent models based on new observational constraints predict mass-loss rates that can be orders of magnitude lower than the canonical prescriptions traditionally used in stellar evolution, but with similar or even higher mass loss than legacy models at log(L/L⊙) > 5.5, i.e. the extrapolation regime past the Humphrey-Davidson limit. They note that the various recipes for cool supergiant winds can be distinguished by their primary physical dependencies and that models from Beasor et al. (2023) and Decin et al. (2024) are parametrized without an explicit effective temperature dependency and have a strong inverse dependency on MZAMS."

Other Uncertainties Discussed: The paper covers additional sources of uncertainty including: Fe and CNO driving limits at low metallicity (noting that below a metallicity of roughly the Small Magellanic Cloud (SMC) value of 0.2 Z⊙, the Fe opacity peak becomes insufficient to drive the wind), envelope inflation, internal mixing and rotation, LBV winds, mass-loss peaks and interpolation/maximization schemes, massive stars in binaries and multiples, magnetic fields, Population II and III stars, and pulsations.

Main Conclusion: The authors conclude that the dominant stellar wind uncertainties arise from a mismatch between the physical assumptions in stellar wind models and the structure of the stars to which they are applied and that the choice of a specific mass-loss recipe is not a minor implementation detail but a dominant source of uncertainty that fundamentally shapes the predicted evolution of massive stars.

Based on this paper, here are specific improvements for AI systems used in astrophysics research:

• Improvement: Implement automatic validation of physical parameters against each prescription's domain of validity (e.g., temperature ranges, metallicity regimes, mass ranges)

• Capability: AI can automatically warn researchers when they apply a wind recipe outside its calibrated regime, preventing invalid extrapolations (e.g., applying Vink et al. 2001 below 25 kK where the bistability jump is disputed, or using Γe-based models below 15 kK where Thomson scattering fails)

• Improvement: Build an ensemble-based system that simultaneously evaluates all 31 wind prescriptions from Table 1 and outputs a probability distribution of mass-loss rates rather than a single value

• Capability: AI can provide researchers with uncertainty ranges (e.g., mass-loss rate spans 10-6 to 10-4 M⊙/yr depending on prescription) and identify where prescriptions diverge most, helping prioritize which physics needs better constraints

• Improvement: Create an AI module that detects when a stellar evolution model uses inconsistent transition criteria (e.g., using Γe-based transitions below 30 kK, or applying chemically homogeneous mass assumptions to non-homogeneous stars)

• Capability: AI can flag when models violate the cool Wolf-Rayet problem constraints, ensuring that thick-wind transitions are only triggered when the underlying physics (electron scattering dominance) is valid

• Improvement: Implement automatic detection and correction of inconsistent solar metallicity (Z⊙) calibrations across combined prescriptions (0.019 vs 0.017 vs 0.0142)

• Capability: AI can standardize or explicitly report which Z⊙ convention each recipe uses, preventing subtle but systematic errors when combining thin-wind and thick-wind recipes in the same model

• Improvement: Train an AI to identify whether a model's mass-loss discontinuity at 25 kK is physically justified or an artifact of outdated assumptions

• Capability: AI can automatically flag models that include the disputed bistability jump and suggest alternatives (e.g., Krtička et al. 2024 or Pauli et al. 2025) that show weaker or no jump, based on recent observational evidence

• Improvement: Implement safeguards for recipes like Beasor et al. (2023) and Decin et al. (2024) that depend on MZAMS, which loses meaning for merged or binary stars

• Capability: AI can detect when a model's mass-loss calculation uses initial mass for stars that have undergone mergers or mass transfer, and recommend switching to current-parameter-based prescriptions

• Improvement: Build an AI classifier that determines which opacity mechanism dominates (Fe, CNO, Si, or Thomson scattering) based on a star's metallicity and effective temperature

• Capability: AI can automatically select the appropriate wind-driving physics for low-metallicity regimes (below 0.2 Z⊙), preventing incorrect Fe-based scaling extrapolations to environments where CNO or Si dominate

• Improvement: Create an AI that checks whether combined wind schemes (e.g., thin + thick + cool supergiant) maintain physical consistency at their boundaries

• Capability: AI can identify discontinuities or artificial jumps at transition temperatures (e.g., 4 kK, 10 kK, 30 kK) and suggest interpolation schemes or flag where maximization approaches (like GENEC's max(V01, GH08)) may overestimate mass loss

• Improvement: Train an AI to cross-reference predicted mass-loss rates against observational datasets (e.g., González-Torà et al. 2023, 2024; Christodoulou et al. 2025)

• Capability: AI can automatically rank prescriptions by their agreement with empirical data for specific stellar types (e.g., RSGs, YSGs, WR stars) and provide confidence scores for each recipe in different parameter regimes

• Improvement: Implement corrections for prescriptions that assume isolated stars, adding terms for tidal stripping, pulsar wind ablation, and merger products

• Capability: AI can adjust mass-loss predictions for binary systems where MZAMS-based recipes (Beasor, Decin) fail, and where companion interactions significantly alter surface properties and binding energy

• Improvement: Track a star's evolutionary history to ensure mass-loss prescriptions are applied consistently across phases (e.g., not switching between different Z⊙ calibrations mid-evolution)

• Capability: AI can detect when a model changes wind schemes without physical justification and flag potential numerical artifacts from discontinuities

• Improvement: Expand the Argus dictionary concept to include automatic cross-referencing of all physical limitations, transition criteria, and calibration ranges from this paper

• Capability: AI can generate complete, physically valid MESA configurations by natural language requests, automatically warning about invalid parameter combinations (e.g., using Γe transition below 15 kK) before simulation starts

Stellar winds are a major source of uncertainty in understanding the life and deaths of massive stars. Across studies in the field, prescriptions for stellar winds differ substantially in both their physical assumptions and implementation, making them a dominant contributor to model-to-model variation. In this work, we present a systematic analysis of the physical assumptions underlying commonly adopted wind prescriptions for optically thin and optically thick winds of hot stars, as well as the winds of cool supergiants. Our analysis reveals substantial discrepancies across all regimes: predicted mass-loss rates for optically thin winds differ by more than an order of magnitude, while rates for cool supergiants vary by several orders of magnitude, with even wider uncertainties arising in extrapolation regimes beyond the Humphreys-Davidson limit. These disparities introduce significant ambiguity into the predicted formation of Wolf-Rayet (WR) stars, a problem further compounded by the inconsistent application of transition criteria. A central issue is the "cool Wolf-Rayet problem", a temperature regime where the classical electron-scattering Eddington factor (e) loses physical consistency. Because this factor is widely used to determine WR mass-loss rates, its failure forces models to rely on uncertain extrapolations and ad-hoc corrections. We conclude that the dominant stellar wind uncertainties arise from a mismatch between the physical assumptions in stellar wind models and the structure of the stars to which they are applied. Our framework clarifies the origins of current theoretical discrepancies and identifies the key physical bottlenecks that must be addressed to improve mass-loss modeling for massive stars.