arXiv subjects

Henry W. Leung

Publications and source records attributed to Henry W. Leung.

12 recordsLinked to original sources

arXiv2024

The Multimodal Universe: Enabling Large-Scale Machine Learning with 100TB of Astronomical Scientific Data

We present the MULTIMODAL UNIVERSE, a large-scale multimodal dataset of scientific astronomical data, compiled specifically to facilitate machine learning research. Overall, the MULTIMODAL UNIVERSE contains hundreds of millions of astronomical observations, constituting 100\,TB of multi-channel and hyper-spectral images, spectra, multivariate time series, as well as a wide variety of associated scientific measurements and "metadata". In addition, we include a range of benchmark tasks representative of standard practices for machine learning methods in astrophysics. This massive dataset will enable the development of large multi-modal models specifically targeted towards scientific applications. All codes used to compile the MULTIMODAL UNIVERSE and a description of how to access the data is available at https://github.com/MultimodalUniverse/MultimodalUniverse

astro-ph.IM↗

arXiv2024

Estimating Probability Densities with Transformer and Denoising Diffusion

Transformers are often the go-to architecture to build foundation models that ingest a large amount of training data. But these models do not estimate the probability density distribution when trained on regression problems, yet obtaining full probabilistic outputs is crucial to many fields of science, where the probability distribution of the answer can be non-Gaussian and multimodal. In this work, we demonstrate that training a probabilistic model using a denoising diffusion head on top of the Transformer provides reasonable probability density estimation even for high-dimensional inputs. The combined Transformer+Denoising Diffusion model allows conditioning the output probability density on arbitrary combinations of inputs and it is thus a highly flexible density function emulator of all possible input/output combinations. We illustrate our Transformer+Denoising Diffusion model by training it on a large dataset of astronomical observations and measured labels of stars within our Galaxy and we apply it to a variety of inference tasks to show that the model can infer labels accurately with reasonable distributions.

cs.LG↗

arXiv2023

Towards an astronomical foundation model for stars with a Transformer-based model

Rapid strides are currently being made in the field of artificial intelligence using Transformer-based models like Large Language Models (LLMs). The potential of these methods for creating a single, large, versatile model in astronomy has not yet been explored. In this work, we propose a framework for data-driven astronomy that uses the same core techniques and architecture as used by LLMs. Using a variety of observations and labels of stars as an example, we build a Transformer-based model and train it in a self-supervised manner with cross-survey data sets to perform a variety of inference tasks. In particular, we demonstrate that a $\textit{single}$ model can perform both discriminative and generative tasks even if the model was not trained or fine-tuned to do any specific task. For example, on the discriminative task of deriving stellar parameters from Gaia XP spectra, we achieve an accuracy of 47 K in $T_\mathrm{eff}$, 0.11 dex in $\log{g}$, and 0.07 dex in $[\mathrm{M/H}]$, outperforming an expert $\texttt{XGBoost}$ model in the same setting. But the same model can also generate XP spectra from stellar parameters, inpaint unobserved spectral regions, extract empirical stellar loci, and even determine the interstellar extinction curve. Our framework demonstrates that building and training a $\textit{single}$ foundation model without fine-tuning using data and parameters from multiple surveys to predict unmeasured observations and parameters is well within reach. Such "Large Astronomy Models" trained on large quantities of observational data will play a large role in the analysis of current and future large surveys.

astro-ph.IM↗

arXiv2023

Decoding the age-chemical structure of the Milky Way disk: An application of Copulas and Elicitable Maps

In the Milky Way, the distribution of stars in the $[α/\mathrm{Fe}]$ vs. $[\mathrm{Fe/H}]$ and $[\mathrm{Fe/H}]$ vs. age planes holds essential information about the history of star formation, accretion, and dynamical evolution of the Galactic disk. We investigate these planes by applying novel statistical methods called copulas and elicitable maps to the ages and abundances of red giants in the APOGEE survey. We find that the low- and high-$α$ disk stars have a clean separation in copula space and use this to provide an automated separation of the $α$ sequences using a purely statistical approach. This separation reveals that the high-$α$ disk ends at the same [$α$/Fe] and age at high $[\mathrm{Fe/H}]$ as the low-$[\mathrm{Fe/H}]$ start of the low-$α$ disk, thus supporting a sequential formation scenario for the high- and low-$α$ disks. We then combine copulas with elicitable maps to precisely obtain the correlation between stellar age $τ$ and metallicity $[\mathrm{Fe/H}]$ conditional on Galactocentric radius $R$ and height $z$ in the range $0 < R < 20$ kpc and $|z| < 2$ kpc. The resulting trends in the age-metallicity correlation with radius, height, and [$α$/Fe] demonstrate a $\approx 0$ correlation wherever kinematically-cold orbits dominate, while the naively-expected negative correlation is present where kinematically-hot orbits dominate. This is consistent with the effects of spiral-driven radial migration, which must be strong enough to completely flatten the age-metallicity structure of the low-$α$ disk.

astro-ph.GA↗

arXiv2023

A variational encoder-decoder approach to precise spectroscopic age estimation for large Galactic surveys

Constraints on the formation and evolution of the Milky Way Galaxy require multi-dimensional measurements of kinematics, abundances, and ages for a large population of stars. Ages for luminous giants, which can be seen to large distances, are an essential component of studies of the Milky Way, but they are traditionally very difficult to estimate precisely for a large dataset and often require careful analysis on a star-by-star basis in asteroseismology. Because spectra are easier to obtain for large samples, being able to determine precise ages from spectra allows for large age samples to be constructed, but spectroscopic ages are often imprecise and contaminated by abundance correlations. Here we present an application of a variational encoder-decoder on cross-domain astronomical data to solve these issues. The model is trained on pairs of observations from APOGEE and Kepler of the same star in order to reduce the dimensionality of the APOGEE spectra in a latent space while removing abundance information. The low dimensional latent representation of these spectra can then be trained to predict age with just $\sim$ 1,000 precise seismic ages. We demonstrate that this model produces more precise spectroscopic ages ($\sim$ 22% overall, $\sim$ 11% for red-clump stars) than previous data-driven spectroscopic ages while being less contaminated by abundance information (in particular, our ages do not depend on [$α$/M]). We create a public age catalog for the APOGEE DR17 data set and use it to map the age distribution and the age-[Fe/H]-[$α$/M] distribution across the radial range of the Galactic disk.

astro-ph.GA↗

arXiv2022

A direct measurement of the distance to the Galactic center using the kinematics of bar stars

The distance to the Galactic center $R_0$ is a fundamental parameter for understanding the Milky Way, because all observations of our Galaxy are made from our heliocentric reference point. The uncertainty in $R_0$ limits our knowledge of many aspects of the Milky Way, including its total mass and the relative mass of its major components, and any orbital parameters of stars employed in chemo-dynamical analyses. While measurements of $R_0$ have been improving over a century, measurements in the past few years from a variety of methods still find a wide range of $R_0$ being somewhere within $8.0$ to $8.5\,\mathrm{kpc}$. The most precise measurements to date have to assume that Sgr A$^*$ is at rest at the Galactic center, which may not be the case. In this paper, we use maps of the kinematics of stars in the Galactic bar derived from APOGEE DR17 and Gaia EDR3 data augmented with spectro-photometric distances from the \texttt{astroNN} neural-network method. These maps clearly display the minimum in the rotational velocity $v_T$ and the quadrupolar signature in radial velocity $v_R$ expected for stars orbiting in a bar. From the minimum in $v_T$, we measure $R_0 = 8.23 \pm 0.12\,\mathrm{kpc}$. We validate our measurement using realistic $N$-body simulations of the Milky Way. We further measure the pattern speed of the bar to be $Ω_\mathrm{bar} = 40.08\pm1.78\,\mathrm{km\,s}^{-1}\mathrm{kpc}^{-1}$. Because the bar forms out of the disk, its center is manifestly the barycenter of the bar+disc system and our measurement is therefore the most robust and accurate measurement of $R_0$ to date.

astro-ph.GA↗

arXiv2021

Chemical Cartography with APOGEE: Mapping Disk Populations with a Two-Process Model and Residual Abundances

We apply a novel statistical analysis to measurements of 16 elemental abundances in 34,410 Milky Way disk stars from the final data release (DR17) of APOGEE-2. Building on recent work, we fit median abundance ratio trends [X/Mg] vs. [Mg/H] with a 2-process model, which decomposes abundance patterns into a "prompt" component tracing core collapse supernovae and a "delayed" component tracing Type Ia supernovae. For each sample star, we fit the amplitudes of these two components, then compute the residuals Δ[X/H] from this two-parameter fit. The rms residuals range from ~0.01-0.03 dex for the most precisely measured APOGEE abundances to ~0.1 dex for Na, V, and Ce. The correlations of residuals reveal a complex underlying structure, including a correlated element group comprised of Ca, Na, Al, K, Cr, and Ce and a separate group comprised of Ni, V, Mn, and Co. Selecting stars poorly fit by the 2-process model reveals a rich variety of physical outliers and sometimes subtle measurement errors. Residual abundances allow comparison of populations controlled for differences in metallicity and [α/Fe]. Relative to the main disk (R=3-13 kpc, |Z|<2 kpc), we find nearly identical abundance patterns in the outer disk (R=15-17 kpc), 0.05-0.2 dex depressions of multiple elements in LMC and Gaia Sausage/Enceladus stars, and wild deviations (0.4-1 dex) of multiple elements in ωCen. Residual abundance analysis opens new opportunities for discovering chemically distinctive stars and stellar populations, for empirically constraining nucleosynthetic yields, and for testing chemical evolution models that include stochasticity in the production and redistribution of elements.

astro-ph.GA↗

arXiv2020

Searching for Solar Siblings in APOGEE and $Gaia$ DR2 with N-body Simulations

We make use of APOGEE and $Gaia$ data to identify stars that are consistent with being born in the same association or star cluster as the Sun. We limit our analysis to stars that match solar abundances within their uncertainties, as they could have formed from the same Giant Molecular Cloud (GMC) as the Sun. We constrain the range of orbital actions that solar siblings can have with a suite of simulations of solar birth clusters evolved in static and time-dependent tidal fields. The static components of each galaxy model are the bulge, disk, and halo, while the various time-dependent components include a bar, spiral arms, and GMCs. In galaxy models without GMCs, simulated solar siblings all have $J_R < 122$ km $\rm s^{-1}$ kpc, $990 < L_z < 1986$ km $\rm s^{-1}$ kpc, and $0.15 < J_z < 0.58$ km $\rm s^{-1}$ kpc. Given the actions of stars in APOGEE and $Gaia$, we find 104 stars that fall within this range. One candidate in particular, Solar Sibling 1, has both chemistry and actions similar enough to the solar values that strong interactions with the bar or spiral arms are not required for it to be dynamically associated with the Sun. Adding GMCs to the potential can eject solar siblings out of the plane of the disk and increase their $J_z$, resulting in a final candidate list of 296 stars. The entire suite of simulations indicate that solar siblings should have $J_R < 122$ km $\rm s^{-1}$ kpc, $353 < L_z < 2110$ km $\rm s^{-1}$ kpc, and $J_z < 0.8$ km $\rm s^{-1}$ kpc. Given these criteria, it is most likely that the association or cluster that the Sun was born in has reached dissolution and is not the commonly cited open cluster M67.

astro-ph.GA↗

arXiv2019

Life in the fast lane: a direct view of the dynamics, formation, and evolution of the Milky Way's bar

Studies of the ages, abundances, and motions of individual stars in the Milky Way provide one of the best ways to study the evolution of disk galaxies over cosmic time. The formation of the Milky Way's barred inner region in particular is a crucial piece of the puzzle of disk galaxy evolution. Using data from APOGEE and Gaia, we present maps of the kinematics, elemental abundances, and age of the Milky Way bulge and disk that show the barred structure of the inner Milky Way in unprecedented detail. The kinematic maps allow a direct, purely kinematic determination of the bar's pattern speed of 41+/-3 km/s/kpc and of its shape and radial profile. We find the bar's age, metallicity, and abundance ratios to be the same as those of the oldest stars in the disk that are formed in its turbulent beginnings, while stars in the bulge outside of the bar are younger and more metal-rich. This implies that the bar likely formed ~8 Gyr ago, when the decrease in turbulence in the gas disk allowed a thin disk to form that quickly became bar-unstable. The bar's formation therefore stands as a crucial epoch in the evolution of the Milky Way, a picture that is in line with the evolutionary path that emerges from observations of the gas kinematics in external disk galaxies over the last ~10 Gyr.

astro-ph.GA↗

arXiv2019

Dynamical heating across the Milky Way disc using APOGEE and $\it{Gaia}$

The kinematics of the Milky Way disc as a function of age are well measured at the solar radius, but have not been studied over a wider range of Galactocentric radii. Here, we measure the kinematics of mono-age, mono-$\mathrm{[Fe/H]}$ populations in the low and high $\mathrm{[α/Fe]}$ discs between $4 \lesssim R \lesssim 13$ kpc and $|z| \lesssim 2$ kpc using 65,719 stars in common between APOGEE DR14 and $\it{Gaia}$ DR2 for which we estimate ages using a Bayesian neural network model trained on asteroseismic ages. We determine the vertical and radial velocity dispersions, finding that the low and high $\mathrm{[α/Fe]}$ discs display markedly different age--velocity-dispersion relations (AVRs) and shapes $σ_z/σ_R$. The high $\mathrm{[α/Fe]}$ disc has roughly flat AVRs and constant $σ_z/σ_R = 0.64\pm 0.04$, whereas the low $\mathrm{[α/Fe]}$ disc has large variations in this ratio which positively correlate with the mean orbital radius of the population at fixed age. The high $\mathrm{[α/Fe]}$ disc component's flat AVRs and constant $σ_z/σ_R$ clearly indicates an entirely different heating history. Outer disc populations also have flatter radial AVRs than those in the inner disc, likely due to the waning effect of spiral arms. Our detailed measurements of AVRs and $σ_z/σ_R$ across the disc indicate that low $\mathrm{[α/Fe]}$, inner disc ($R \lesssim 10\,\mathrm{kpc}$) stellar populations are likely dynamically heated by both giant molecular clouds and spiral arms, while the observed trends for outer disc populations require a significant contribution from another heating mechanism such as satellite perturbations. We also find that outer disc populations have slightly positive mean vertical and radial velocities, likely because they are part of the warped disc.

astro-ph.GA↗

arXiv2019

Simultaneous calibration of spectro-photometric distances and the Gaia DR2 parallax zero-point offset with deep learning

Gaia measures the five astrometric parameters for stars in the Milky Way, but only four of them (positions and proper motion, but not parallax) are well measured beyond a few kpc from the Sun. Modern spectroscopic surveys such as APOGEE cover a large area of the Milky Way disk and we can use the relation between spectra and luminosity to determine distances to stars beyond Gaia's parallax reach. Here, we design a deep neural network trained on stars in common between Gaia and APOGEE that determines spectro-photometric distances to APOGEE stars, while including a flexible model to calibrate parallax zero-point biases in Gaia DR2. We determine the zero-point offset to be $-52.3 \pm 2.0uas$ when modeling it as a global constant, but also train a multivariate zero-point offset model that depends on $G$, $G_{BP} - G_{RP}$ color, and $T_\mathrm{eff}$ and that can be applied to all 139 million stars in Gaia DR2 within APOGEE's color--magnitude range. Our spectro-photometric distances are more precise than Gaia at distances $\approx 2kpc$ from the Sun. We release a catalog of spectro-photometric distances for the entire APOGEE DR14 data set which covers Galactocentric radii $2kpc\lesssim R \lesssim19kpc$; $\approx 150,000$ stars have <10% uncertainty, making this a powerful sample to study the chemo-dynamical structure of the disk. We use this sample to map the mean [Fe/H] and 15 abundance ratios [X/Fe] from the Galactic center to the edge of the disk. Among many interesting trends, we find that the bulge and bar region at $R \lesssim 5kpc$ clearly stands out in [Fe/H] and most abundance ratios.

astro-ph.GA↗

arXiv2019

Deep learning of multi-element abundances from high-resolution spectroscopic data

Deep learning with artificial neural networks is increasingly gaining attention, because of its potential for data-driven astronomy. However, this methodology usually does not provide uncertainties and does not deal with incompleteness and noise in the training data. In this work, we design a neural network for high-resolution spectroscopic analysis using APOGEE data that mimics the methodology of standard spectroscopic analyses: stellar parameters are determined using the full wavelength range, but individual element abundances use censored portions of the spectrum. We train this network with a customized objective function that deals with incomplete and noisy training data and apply dropout variational inference to derive uncertainties on our predictions. We determine parameters and abundances for 18 individual elements at the $\approx 0.03$ dex level, even at low signal-to-noise ratio. We demonstrate that the uncertainties returned by our method are a realistic estimate of the precision and they automatically blow up when inputs or outputs outside of the training set are encountered, thus shielding users from unwanted extrapolation. By using standard deep-learning tools for GPU acceleration, our method is extremely fast, allowing analysis of the entire APOGEE data set of $\approx250,000$ spectra in ten minutes on a single, low-cost GPU. We release the stellar parameters and 18 individual-element abundances with associated uncertainty for the entire APOGEE DR14 dataset. Simultaneously, we release astroNN, a well-tested, open-source python package developed for this work, but that is also designed to be a general package for deep learning in astronomy. astroNN is available at https://github.com/henrysky/astroNN with extensive documentation at http://astroNN.readthedocs.io.

astro-ph.GA↗