Machine learning connects atomic physics to simulations of stellar winds

Light can push gas away from a star, but calculating that push requires more than knowing how brightly the star shines. The force also depends on which wavelengths reach the gas and which atomic transitions can absorb them. As radiation travels through an outflow, absorption changes its spectrum, potentially changing the wind itself.

In a preprint, Sergei Dyda of the University of Alabama and Randall C. Dannen of the University of Nevada, Las Vegas present a machine-learning method for connecting those atomic processes to simulations of moving gas. Their approach learns from detailed photoionization calculations, then supplies estimates of the radiation force during a simulation. The work is a computational demonstration, not an observational discovery; the supplied preprint does not establish that it has undergone peer review.

Learning how light pushes gas

The central quantity is a “force multiplier,” which expresses the strength of radiation’s push through atomic spectral lines relative to electron scattering. Calculating it requires information about the gas’s ionization state and the opacity of its many atomic transitions.

Running a full photoionization calculation at every location and timestep would be computationally expensive. Precomputed tables offer an alternative, but become harder to construct and use as more variables are added. The authors instead train a surrogate: a machine-learning model designed to reproduce selected outputs of the detailed calculation, rather than solve the atomic physics anew.

They generated training data with XSTAR, using a version with more than two million spectral lines. Their dataset contained 177,100 samples spanning three inputs: an ionization parameter, an optical-depth parameter, and the temperature describing an illuminating blackbody spectrum. Blackbody temperatures ranged from 10,000 to one million kelvin.

The researchers randomly reserved 20% of the samples—35,420 points—for testing. Their baseline model used 80,000 training points and predicted force multipliers in four bands: infrared, optical, ultraviolet and X-rays.

A stronger result than table interpolation

Among the configurations tested, a random forest performed better than gradient boosting and a multilayer neural network. The adopted forest combined 200 decision trees. That ranking applies to this dataset and these implementations, not to machine-learning methods universally.

The authors assessed errors logarithmically. A difference of 0.01 dex corresponds to about 2.3%, while 0.05 dex corresponds to roughly 12%. With 80,000 training points, around 60% of predictions met the tighter accuracy threshold. Larger training sets improved performance.

A separate comparison tested the forest against trilinear interpolation, which estimates a value using eight neighboring grid points. In that benchmark, 1.2% of machine-learning predictions had errors above approximately 0.05 dex, compared with more than 40% for interpolation.

Those figures have important boundaries. The comparison excluded points at the edges of the parameter space and cases where the total force multiplier was below 0.01, which the authors considered dynamically negligible. The random test split also assessed prediction within the sampled family, not performance on wholly different radiation spectra.

Putting the model inside a wind

The team exported the trained model into Athena++, a hydrodynamics code. During a simulation, the code supplied local conditions to the forest and used its predicted force multipliers to calculate line acceleration.

The demonstration followed spherically symmetric winds from an idealized star with 50 times the Sun’s mass and 10 times its radius. The authors explicitly describe this as a controlled test, not a detailed representation of a particular observed star.

They compared winds with an unchanged spectral shape against winds in which preferential absorption of ultraviolet or X-ray radiation altered the spectrum with radius. The resulting mass flow and velocity depended on that spectral evolution. In the X-ray-absorption case, the mass flux matched a comparison model with a constant blackbody temperature of 400,000 kelvin, while the outflow velocity increased by about 20%.

What remains approximate

Absorption generally turns an initially blackbody spectrum into something that is no longer a blackbody. Here, the simulation represented the altered spectrum with an equivalent blackbody temperature chosen from its ultraviolet flux fraction. That preserves one measure of spectral shape, but cannot capture every distortion.

The authors also caution against using the model outside its training range. Extending it to more complex radiation fields would require suitable spectral inputs and additional training data. The paper describes rapid runtime evaluation, but does not report a quantified end-to-end simulation speedup.

The demonstrated advance is therefore specific: a learned approximation can supply radiation-force information inside a working wind simulation. Its usefulness for more realistic systems will depend on whether the training data and spectral description capture the physics those systems require.