Periodic Reporting for period 3 - CartographY (Mapping Stellar Helium)
Reporting period: 2022-04-01 to 2023-09-30
Much of modern-day astrophysics is underpinned by the assumption that we have reliable models for stellar evolution in order to accurately ‘model’ stars. Unfortunately we have very little to prove this is the case. A well-known problem is that we rarely have a complete set of observations of any given star - from a small number of observations we must piece together the true picture. But, whatever observations we have, we do not have a direct measure of the initial stellar helium content. Initial He has a dramatic impact on inferred masses and ages of stars - e.g. an increase in initial He fraction of 0.05 for a Sun-like star will lead to a reduction of stellar MS lifetime of ~50%. Hence any quoted age is only valid for the adopted treatment of initial He and uncertainties quoted are a measure of precision not accuracy. This is not good enough for the many fields of astrophysics that rely on accurate stellar masses and ages.
In most real world cases progress has been made by tying initial He to initial metallicity but these relations are at best poorly calibrated, or at worst completely inappropriate. This systematic failure limits our understanding of our universe when we try to understand stars, the demographics of exoplanets, and the structure and evolution of our own Milky Way.
In most real world cases progress has been made by tying initial He to initial metallicity but these relations are at best poorly calibrated, or at worst completely inappropriate. This systematic failure limits our understanding of our universe when we try to understand stars, the demographics of exoplanets, and the structure and evolution of our own Milky Way.
To face this challenge we have used a data-driven approach, using Bayesian Hierarchical Models (BHM’s), to leverage the exceptionally diverse and numerous observations of stars to solve the problem of initial He. BHM’s allow us to beat the curse of dimensionality that would otherwise render this task computationally impossible. By using multi-level BHM’s we will use the statistics of the population to map out the initial He of stars and then generate precise and accurate estimates of fundamental stellar parameters.
We have addressed a number of challenges of applying BHM's to the problem of stellar evolution. By far the most challenging aspect of the problem, is that using, models of stellar evolution is slow - very slow. We have used machine learning to speed up this step by producing a neural network emulator of stellar evolution. This step has reduced out computational time scale by 6 orders of magnitude and given us access to mathematical properties of our models that would otherwise be impossible to calculate.
By combining our neural network emulator with state-of-the-art probabilistic programming languages we have been able to simultaneously model the population and the individual properties of 80 stars. And this method we have scales well. Tests of artificial data suggest we could now deploy our method to over tern thousand stars with a run time of less than a single day.
We have addressed a number of challenges of applying BHM's to the problem of stellar evolution. By far the most challenging aspect of the problem, is that using, models of stellar evolution is slow - very slow. We have used machine learning to speed up this step by producing a neural network emulator of stellar evolution. This step has reduced out computational time scale by 6 orders of magnitude and given us access to mathematical properties of our models that would otherwise be impossible to calculate.
By combining our neural network emulator with state-of-the-art probabilistic programming languages we have been able to simultaneously model the population and the individual properties of 80 stars. And this method we have scales well. Tests of artificial data suggest we could now deploy our method to over tern thousand stars with a run time of less than a single day.
Our application has already led to much better constraint on stellar fundamental parameters and out constraint of the enrichment of Helium in our stellar neighbourhood. This constraint will only improve as we increase the number of stars we can process into the hundreds of thousands.