kmaltsev/stellar-evolution-emulators: Stellar-evolution-forecasting

This release contains a machine-learning model that efficiently predicts the classical photometric observables 

• bolometric luminosity $\log L/L_\odot$, 
• effective temperature $\log T_\mathrm{eff}/\mathrm{K}$ and 
•  surface gravity $\log g/\mathrm{[cm/s^2]}$ 

during stellar evolution from the zero-age-main-sequence (ZAMS) up to the end of core-helium burning over a ZAMS mass range $M_\mathrm{ZAMS}/M_\odot \in (0.7, 300)$ at solar metallicity $Z=Z_\odot$. On an ordinary 4-core CPU, it takes about 40 seconds to cast 1 million point predictions, and the predictive error is - depending on the observable - 1-3 orders of magnitude lower than typical observational uncertainties. Since the model covers >99% of stellar lifetimes, it can be used for comparison to observations of an individual star or stellar populations at $Z=Z_\odot$, for example,

•  to estimate ZAMS mass of each star and infer the Initial-Mass-Function (IMF) of the population, or
•  to estimate the age of each star, or
•  to test the adopted physics underlying this stellar evolution model by assessing how well it reproduces the observations.

In the Jupyter Notebook stellar-evolution-emulator-fitted-models.ipynb, it is shown how to use this machine-learning model based on a few simple test examples.

The HNNI.ipynb Notebook contains the Hierarchical Nearest-Neighbor Interpolation (HNNI) algorithm, which is an alternative method for automated interpolation of stellar evolution tracks from the ZAMS up to the end of core-helium burning over $M_\mathrm{ZAMS}/M_\odot \in (0.7, 300)$ at $Z=Z_\odot$. HNNI is more accurate than the machine-learning based surrogate model, and predicts not only the observables but any stellar evolution variable of interest. However, it requires continued access to the underlying stellar evolution catalog data, and - on a 4-core CPU - requires about 1.5 hours to cast 1 million point predictions. While HNNI has been demonstrated to work on the MIST data set, it can be used to interpolate any stellar evolution catalog of interest. Contrary to other stellar track interpolation methods, such as those used e.g. in context of population synthesis, HNNI does not require the segmentation of the catalog data into separate evolutionary phases in order to interpolate stellar variables during the entire evolutionary sequence from ZAMS up to the end of core-helium burning. 

Both methods have in common that, instead of stellar age, a timescale-adapted evolutionary coordinate $s$ is used to parametrize the evolution of stars. When a star is at ZAMS, $s=0$, and when it terminates core-helium burning, $s=1$. The advantage of $s$ is that 

• it reduces timescale variability when visualizing and fitting stellar variables during the evolutionary sequence across timescale-separated evolutionary phases, and that
• it allows to quickly produce stellar tracks e.g. in the Hertzsprung-Russell and Kiel diagrams without compromising the resolution of fast-timescale evolutionary sequences (e.g. the Hertzsprung gap).  

Both methods are developed in Maltsev et al. 2024, based on the MIST stellar evolution catalog pre-computed by Choi et al. 2016.