Machine Learning the Sixth Dimension: Stellar Radial Velocities from 5D Phase-space Correlations

Dropulic, A.; Ostdiek, Bryan; Chang, Luke L. Y.; Liu, Hongwan; Cohen, Timothy; Lisanti, Mariangela

doi:10.3847/2041-8213/ac09ef

Cited by 12 publications

(9 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…their proper motions), we might attain better results by instead retaining these stars and estimating their missing los velocities. A possible technique to do so has been suggested by the work of Dropulic et al ( 2021 ), which demonstrated that artificial neural networks can be successful in recreating the missing los velocities of Gaia stars.…”

Section: The Local Acceleration Fieldmentioning

confidence: 99%

Charting galactic accelerations – II. How to ‘learn’ accelerations in the solar neighbourhood

Naik

Burrage

et al. 2022

Monthly Notices of the Royal Astronomical Society

View full text Add to dashboard Cite

Gravitational acceleration fields can be deduced from the collisionless Boltzmann equation, once the distribution function is known. This can be constructed via the method of normalizing flows from datasets of the positions and velocities of stars. Here, we consider application of this technique to the solar neighbourhood. We construct mock data from a linear superposition of multiple ‘quasi-isothermal’ distribution functions, representing stellar populations in the equilibrium Milky Way disc. We show that given a mock dataset comprising a million stars within 1 kpc of the Sun, the underlying acceleration field can be measured with excellent, sub-percent level accuracy, even in the face of realistic errors and missing line-of-sight velocities. The effects of disequilibrium can lead to bias in the inferred acceleration field. This can be diagnosed by the presence of a phase space spiral, which can be extracted simply and cleanly from the learned distribution function. We carry out a comparison with two other popular methods of finding the local acceleration field (Jeans analysis and 1D distribution function fitting). We show our method most accurately measures accelerations from a given mock dataset, particularly in the presence of disequilibria.

show abstract

Section: The Local Acceleration Fieldmentioning

confidence: 99%

Charting galactic accelerations – II. How to ‘learn’ accelerations in the solar neighbourhood

Naik

Burrage

et al. 2022

Monthly Notices of the Royal Astronomical Society

View full text Add to dashboard Cite

show abstract

“…Given that the discarded stars do have two dimensions of velocity information (i.e., their proper motions), we might attain better results by instead retaining these stars and estimating their missing los velocities. A possible technique to do so has been suggested by the work of Dropulic et al (2021), which demonstrated that artificial neural networks can be successful in recreating the missing los velocities of Gaia stars.…”

Section: The Local Acceleration Fieldmentioning

confidence: 99%

Charting galactic accelerations II: how to 'learn' accelerations in the solar neighbourhood

Naik,

An,

Burrage

et al. 2021

Preprint

View full text Add to dashboard Cite

Gravitational acceleration fields can be deduced from the collisionless Boltzmann equation, once the distribution function is known. This can be constructed via the method of normalizing flows from datasets of the positions and velocities of stars. Here, we consider application of this technique to the solar neighbourhood. We construct mock data from a linear superposition of multiple 'quasi-isothermal' distribution functions, representing stellar populations in the equilibrium Milky Way disc. We show that given a mock dataset comprising a million stars within 1 kpc of the Sun, the underlying acceleration field can be measured with excellent, sub-percent level accuracy, even in the face of realistic errors and missing line-of-sight velocities. The effects of disequilibrium can lead to bias in the inferred acceleration field. This can be diagnosed by the presence of a phase space spiral, which can be extracted simply and cleanly from the learned distribution function. We carry out a comparison with two other popular methods of finding the local acceleration field (Jeans analysis and 1D distribution function fitting). We show our method most accurately measures accelerations from a given mock dataset, particularly in the presence of disequilibria.

show abstract

“…The usage of emulators in astrophysics, based on machine learning (ML) and deep learning techniques, has steadily increased its importance in recent years (LeCun et al 2015). However, so far ML applications in astrophysics are mainly limited to data driven inferences as parameters or classifications; for example Ucci et al (2018) use decision trees to infer key ISM physical properties from emission line ratios, Chardin et al (2019) emulates radiative transfer calculation using Epoch of Reionization simulations as a training dataset, Prelogović et al (2022) infer astrophysical parameters from 21 cm light cone images by adopting recurrent neural networks, and Dropulic et al (2021) shows how to predict stellar line-of-sight velocity from Gaia observations of the Milky Way (MW) by training on phase space mock data sets. Currently, there are few attempts to alleviate the cost of computing chemistry by using auto-encoders: Grassi et al (2021) tries to reduce the complexity of chemical ODE with high dimensionality by compression in a latent space of a smaller dimension; while Grassi et al (2021) showcased the approach for isothermal models, its generalization seems non-trivial.…”

Section: Introductionmentioning

confidence: 99%

Neural networks: solving the chemistry of the interstellar medium

Branca

Pallottini

2022

Monthly Notices of the Royal Astronomical Society

View full text Add to dashboard Cite

Non-equilibrium chemistry is a key process in the study of the InterStellar Medium (ISM), in particular the formation of molecular clouds and thus stars. However, computationally it is among the most difficult tasks to include in astrophysical simulations, because of the typically high (>40) number of reactions, the short evolutionary timescales (about 104 times less than the ISM dynamical time) and the characteristic non-linearity and stiffness of the associated Ordinary Differential Equations system (ODEs). In this proof of concept work, we show that Physics Informed Neural Networks (PINN) are a viable alternative to traditional ODE time integrators for stiff thermo-chemical systems, i.e. up to molecular hydrogen formation (9 species and 46 reactions). Testing different chemical networks in a wide range of densities (−2 < log n/cm−3 < 3) and temperatures (1 < log T/K < 5), we find that a basic architecture can give a comfortable convergence only for simplified chemical systems: to properly capture the sudden chemical and thermal variations a Deep Galerkin Method is needed. Once trained (∼103 GPUhr), the PINN well reproduces the strong non-linear nature of the solutions (errors $\lesssim 10{{\%}}$) and can give speed-ups up to a factor of ∼200 with respect to traditional ODE solvers. Further, the latter have completion times that vary by about $\sim 30{{\%}}$ for different initial n and T, while the PINN method gives negligible variations. Both the speed-up and the potential improvement in load balancing imply that PINN-powered simulations are a very palatable way to solve complex chemical calculation in astrophysical and cosmological problems.

show abstract

Machine Learning the Sixth Dimension: Stellar Radial Velocities from 5D Phase-space Correlations

Cited by 12 publications

References 33 publications

Charting galactic accelerations – II. How to ‘learn’ accelerations in the solar neighbourhood

Charting galactic accelerations – II. How to ‘learn’ accelerations in the solar neighbourhood

Charting galactic accelerations II: how to 'learn' accelerations in the solar neighbourhood

Neural networks: solving the chemistry of the interstellar medium

Contact Info

Product

Resources

About