arXiv subjects

Milos Milosavljevic

Publications and source records attributed to Milos Milosavljevic.

At least 19 recordsLinked to original sources

arXiv2022

Collapsible Linear Blocks for Super-Efficient Super Resolution

With the advent of smart devices that support 4K and 8K resolution, Single Image Super Resolution (SISR) has become an important computer vision problem. However, most super resolution deep networks are computationally very expensive. In this paper, we propose Super-Efficient Super Resolution (SESR) networks that establish a new state-of-the-art for efficient super resolution. Our approach is based on linear overparameterization of CNNs and creates an efficient model architecture for SISR. With theoretical analysis, we uncover the limitations of existing overparameterization methods and show how the proposed method alleviates them. Detailed experiments across six benchmark datasets demonstrate that SESR achieves similar or better image quality than state-of-the-art models while requiring 2x to 330x fewer Multiply-Accumulate (MAC) operations. As a result, SESR can be used on constrained hardware to perform x2 (1080p to 4K) and x4 (1080p to 8K) SISR. Towards this, we estimate hardware performance numbers for a commercial Arm mobile-Neural Processing Unit (NPU) for 1080p to 4K (x2) and 1080p to 8K (x4) SISR. Our results highlight the challenges faced by super resolution on AI accelerators and demonstrate that SESR is significantly faster (e.g., 6x-8x higher FPS) than existing models on mobile-NPU. Finally, SESR outperforms prior models by 1.5x-2x in latency on Arm CPU and GPU when deployed on a real mobile device. The code for this work is available at https://github.com/ARM-software/sesr.

eess.IV↗

arXiv2019

Solving Irregular and Data-enriched Differential Equations using Deep Neural Networks

Recent work has introduced a simple numerical method for solving partial differential equations (PDEs) with deep neural networks (DNNs). This paper reviews and extends the method while applying it to analyze one of the most fundamental features in numerical PDEs and nonlinear analysis: irregular solutions. First, the Sod shock tube solution to compressible Euler equations is discussed, analyzed, and then compared to conventional finite element and finite volume methods. These methods are extended to consider performance improvements and simultaneous parameter space exploration. Next, a shock solution to compressible magnetohydrodynamics (MHD) is solved for, and used in a scenario where experimental data is utilized to enhance a PDE system that is \emph{a priori} insufficient to validate against the observed/experimental data. This is accomplished by enriching the model PDE system with source terms and using supervised training on synthetic experimental data. The resulting DNN framework for PDEs seems to demonstrate almost fantastical ease of system prototyping, natural integration of large data sets (be they synthetic or experimental), all while simultaneously enabling single-pass exploration of the entire parameter space.

cs.LG↗

arXiv2019

Constraining Cosmology with Big Data Statistics of Cosmological Graphs

By utilizing large-scale graph analytic tools implemented in the modern Big Data platform, Apache Spark, we investigate the topological structure of gravitational clustering in five different universes produced by cosmological $N$-body simulations with varying parameters: (1) a WMAP 5-year compatible $Λ$CDM cosmology, (2) two different dark energy equation of state variants, and (3) two different cosmic matter density variants. For the Big Data calculations, we use a custom build of stand-alone Spark/Hadoop cluster at Korea Institute for Advanced Study (KIAS) and Dataproc Compute Engine in Google Cloud Platform (GCP) with the sample size ranging from 7 millions to 200 millions. We find that among the many possible graph-topological measures, three simple ones: (1) the average of number of neighbors (the so-called average vertex degree) $α$, (2) closed-to-connected triple fraction (the so-called transitivity) $τ_Δ$, and (3) the cumulative number density $n_{s\ge5}$ of subcomponents with connected component size $s \ge 5$, can effectively discriminate among the five model universes. Since these graph-topological measures are in direct relation with the usual $n$-points correlation functions of the cosmic density field, graph-topological statistics powered by Big Data computational infrastructure opens a new, intuitive, and computationally efficient window into the dark Universe.

astro-ph.CO↗

arXiv2018

Blind nucleosynthetic source discovery in astronomical elemental abundance data

We demonstrate physics-blind and purely data-based nucleosynthetic archetype discovery from heterogeneous stellar elemental abundance measurements. Archetype extraction is performed via non-negative matrix factorization (NMF). Because stellar elemental abundance measurements can be incomplete and sparsely sampled, we perform combined low-rank NMF and matrix completion on the Hypatia Catalog and JINAbase to extract archetypal elemental abundance patterns, the so-called endmembers, and impute the missing measurements. We qualitatively compare the discovered nucleosynthetic archetypes to patterns anticipated from direct astrophysical nucleosynthetic calculations. Our tentative results indicate that the physics-blind analysis relying entirely on the incomplete chemical mixing in star formation can discern nucleosynthetic archetypes associated with asymptotic giant branch (AGB) stellar sources of the s-process, thermonuclear supernovae, and non-thermonuclear supernovae with an r-process.

astro-ph.SR↗

arXiv2018

Molecular nucleation theory of dust formation in core-collapse supernovae applied to SN 1987A

We model dust formation in the core collapse supernova explosion SN 1987A by treating the gas-phase formation of dust grain nuclei as a chemical process. To compute the synthesis of fourteen species of grains we integrate a non-equilibrium network of nucleating and related chemical reactions and follow the growth of the nuclei into grains via accretion and coagulation. The effects of the radioactive cobalt, titanium, and sodium on the thermodynamics and chemistry of the ejecta are taken into account. The grain temperature, which we allow to differ from the gas temperature, affects the surface-tension-corrected evaporation rate. We also account for He$^+$, Ne$^+$, Ar$^+$, and O weathering. We combine our dust synthesis model with a crude prescription for anisotropic radioactive nickel dredge-up into the core ejecta, the so-called `nickel bubbles', to compute the total dust mass and molecular-species-specific grain size distribution. The total mass varies between $0.41\,M_\odot$ and $0.73\,M_\odot$, depending on the bubble shell density contrast. In the decreasing order of abundance, the grain species produced are: magnesia, silicon, forsterite, iron sulfide, carbon, silicon dioxide, alumina, and iron. The combined grain size distribution is a power law $dN/da\propto a^{-4.39}$. Early ejecta compaction by expanding radioactive nickel bubbles strongly enhances dust synthesis. This underscores the need for improved understanding of hydrodynamic transport and mixing over the entire pre-homologous expansion.

astro-ph.SR↗

arXiv2018

Discrete diffusion Lyman-alpha radiative transfer

Due to its accuracy and generality, Monte Carlo radiative transfer (MCRT) has emerged as the prevalent method for Ly$α$ radiative transfer in arbitrary geometries. The standard MCRT encounters a significant efficiency barrier in the high optical depth, diffusion regime. Multiple acceleration schemes have been developed to improve the efficiency of MCRT but the noise from photon packet discretization remains a challenge. The discrete diffusion Monte Carlo (DDMC) scheme has been successfully applied in state-of-the-art radiation hydrodynamics (RHD) simulations. Still, the established framework is not optimal for resonant line transfer. Inspired by the DDMC paradigm, we present a novel extension to resonant DDMC (rDDMC) in which diffusion in space and frequency are treated on equal footing. We explore the robustness of our new method and demonstrate a level of performance that justifies incorporating the method into existing Ly$α$ codes. We present computational speedups of $\sim 10^2$-$10^6$ relative to contemporary MCRT implementations with schemes that skip scattering in the core of the line profile. This is because the rDDMC runtime scales with the spatial and frequency resolution rather than the number of scatterings - the latter is typically $\propto τ_0$ for static media, or $\propto (a τ_0)^{2/3}$ with core-skipping. We anticipate new frontiers in which on-the-fly Ly$α$ radiative transfer calculations are feasible in 3D RHD. More generally, rDDMC is transferable to any computationally demanding problem amenable to a Fokker-Planck approximation of frequency redistribution.

astro-ph.CO↗

arXiv2017

Radiation pressure in super star cluster formation

The physics of star formation at its extreme, in the nuclei of the densest and the most massive star clusters in the universe - potential massive black hole nurseries - has for decades eluded scrutiny. Spectroscopy of these systems has been scarce, whereas theoretical arguments suggest that radiation pressure on dust grains somehow inhibits star formation. Here, we harness an accelerated Monte Carlo radiation transport scheme to report a radiation hydrodynamical simulation of super star cluster formation in turbulent clouds. We find that radiation pressure reduces the global star formation efficiency by 30-35%, and the star formation rate by 15-50%, both relative to a radiation-free control run. Overall, radiation pressure does not terminate the gas supply for star formation and the final stellar mass of the most massive cluster is $\sim 1.3\times10^6\,M_\odot$. The limited impact as compared to in idealized theoretical models is attributed to a radiation-matter anti-correlation in the supersonically turbulent, gravitationally collapsing medium. In isolated regions outside massive clusters, where the gas distribution is less disturbed, radiation pressure is more effective in limiting star formation. The resulting stellar density at the cluster core is $\ge 10^{8}\,M_{\odot}\,\textrm{pc}^{-3}$, with stellar velocity dispersion $\gtrsim 70\,\text{km}\,\text{s}^{-1}$. We conclude that the super star cluster nucleus is propitious to the formation of very massive stars via dynamical core collapse and stellar merging. We speculate that the very massive star may avoid the claimed catastrophic mass loss by continuing to accrete dense gas condensing from a gravitationally-confined ionized phase.

astro-ph.GA↗

arXiv2016

Toward $ab\,initio$ extremely metal poor stars

Extremely metal poor stars have been the focus of much recent attention owing to the expectation that their chemical abundances can shed light on the metal and dust yields of the earliest supernovae. We present our most realistic simulation to date of the astrophysical pathway to the first metal enriched stars. We simulate the radiative and supernova hydrodynamic feedback of a $60\,M_\odot$ Population III star starting from cosmological initial conditions realizing Gaussian density fluctuations. We follow the gravitational hydrodynamics of the supernova remnant at high spatial resolution through its freely-expanding, adiabatic, and radiative phases, until gas, now metal-enriched, has resumed runaway gravitational collapse. Our findings are surprising: while the Population III progenitor exploded with a low energy of $10^{51}\,\text{erg}$ and injected an ample metal mass of $6\,M_\odot$, the first cloud to collapse after the supernova explosion is a dense surviving primordial cloud on which the supernova blastwave deposited metals only superficially, in a thin, unresolved layer. The first metal-enriched stars can form at a very low metallicity, of only $2-5\times10^{-4}\,Z_\odot$, and can inherit the parent cloud's highly elliptical, radially extended orbit in the dark matter gravitational potential.

astro-ph.GA↗

arXiv2015

An Increasing Stellar Baryon Fraction in Bright Galaxies at High Redshift

Recent observations have shown that the characteristic luminosity of the rest-frame ultraviolet (UV) luminosity function does not significantly evolve at 4 < z < 7 and is approximately M*_UV ~ -21. We investigate this apparent non-evolution by examining a sample of 178 bright, M_UV < -21 galaxies at z=4 to 7, analyzing their stellar populations and host halo masses. Including deep Spitzer/IRAC imaging to constrain the rest-frame optical light, we find that M*_UV galaxies at z=4-7 have similar stellar masses of log(M/Msol)=9.6-9.9 and are thus relatively massive for these high redshifts. However, bright galaxies at z=4-7 are less massive and have younger inferred ages than similarly bright galaxies at z=2-3, even though the two populations have similar star formation rates and levels of dust attenuation. We match the abundances of these bright z=4-7 galaxies to halo mass functions from the Bolshoi Lambda-CDM simulation to estimate the halo masses. We find that the typical halo masses in ~M*_UV galaxies decrease from log(M_h/Msol)=11.9 at z=4 to log(M_h/Msol)=11.4 at z=7. Thus, although we are studying galaxies at a similar mass across multiple redshifts, these galaxies live in lower mass halos at higher redshift. The stellar baryon fraction in units of the cosmic mean Omega_b/Omega_m rises from 5.1% at z=4 to 11.7% at z=7; this evolution is significant at the ~3-sigma level. This rise does not agree with simple expectations of how galaxies grow, and implies that some effect, perhaps a diminishing efficiency of feedback, is allowing a higher fraction of available baryons to be converted into stars at high redshifts.

astro-ph.GA↗

arXiv2015

Star Formation for Predictive Primordial Galaxy Formation

The elegance of inflationary cosmology and cosmological perturbation theory ends with the formation of the first stars and galaxies, the initial sources of light that launched the phenomenologically rich process of cosmic reionization. Here we review the current understanding of early star formation, emphasizing unsolved problems and technical challenges. We begin with the first generation of stars to form after the Big Bang and trace how they influenced subsequent star formation. The onset of chemical enrichment coincided with a sharp increase in the overall physical complexity of star forming systems. Ab-initio computational treatments are just now entering the domain of the predictive and are establishing contact with local observations of the relics of this ancient epoch.

astro-ph.GA↗

arXiv2015

The first galaxies: simulating their feedback-regulated assembly

We investigate the formation of a galaxy reaching a virial mass of $~ 10^8$ solar mass at $z=10$ by carrying out a zoomed radiation-hydrodynamical cosmological simulation. This simulation traces Population~III (Pop~III) star formation, characterized by a modestly top-heavy initial mass function (IMF), and considers stellar feedback such as photoionization heating from Pop III and Population~II (Pop~II) stars, mechanical and chemical feedback from supernovae (SNe), and X-ray feedback from accreting black holes (BHs) and high-mass X-ray binaries (HMXBs). We self-consistently impose a transition in star formation mode from top-heavy Pop III to low-mass Pop~II, and find that the star formation rate in the computational box is dominated by Pop~III until $z=13$, and by Pop~II thereafter. The simulated galaxy experiences bursty star formation, with a substantially reduced gas content due to photoionization heating from Pop~III and Pop~II stars, together with SN feedback. All the gas within the simulated galaxy is metal-enriched above $10^{-5}$ solar, such that there are no remaining pockets of primordial gas. The simulated galaxy has an estimated observed flux of $~10^{-3} nJy$, which is too low to be detected by the James Webb Space Telescope (JWST) without strong lensing amplification. We also show that our simulated galaxy is similar in terms of stellar mass to Segue 2, the least luminous dwarf known in the Local Group.

astro-ph.GA↗

arXiv2015

Radiation pressure driving of a dusty atmosphere

Radiation pressure can be dynamically important in star-forming environments such as ultra-luminous infrared and submillimeter galaxies. Whether and how radiation drives turbulence and bulk outflows in star formation sites is still unclear. The uncertainty in part reflects the limitations of direct numerical schemes that are currently used to simulate radiation transfer and radiation-gas coupling. An idealized setup in which radiation is introduced at the base of a dusty atmosphere in a gravitational field has recently become the standard test for radiation-hydrodynamics methods in the context of star formation. To a series of treatments featuring the flux-limited-diffusion approximation as well as a short-characteristics tracing and M1 closure for the variable Eddington tensor approximation, we here add another treatment that is based on the Implicit Monte Carlo radiation transfer scheme. Consistent with all previous treatments, the atmosphere undergoes Rayleigh-Taylor instability and readjusts to a near-Eddington-limited state. We detect late-time net acceleration in which the turbulent velocity dispersion matches that reported previously with the short-characteristics-based radiation transport closure, the most accurate of the three preceding treatments. Our technical result demonstrates the importance of accurate radiation transfer in simulations of radiative feedback.

astro-ph.GA↗

arXiv2015

Abundance anomalies in metal-poor stars from Population III supernova ejecta hydrodynamics

We present a simulation of the long-term evolution of a Population III supernova remnant in a cosmological minihalo. Employing passive Lagrangian tracer particles, we investigate how chemical stratification and anisotropy in the explosion can affect the abundances of the first low-mass, metal-enriched stars. We find that reverse shock heating can leave the inner mass shells at entropies too high to cool, leading to carbon-enhancement in the re-collapsing gas. This hydrodynamic selection effect could explain the observed incidence of carbon-enhanced metal-poor (CEMP) stars at low metallicity. We further explore how anisotropic ejecta distributions, recently seen in direct numerical simulations of core-collapse explosions, may translate to abundances in metal-poor stars. We find that some of the observed scatter in the Population II abundance ratios can be explained by an incomplete mixing of supernova ejecta, even in the case of only one contributing enrichment event. We demonstrate that the customary hypothesis of fully-mixed ejecta clearly fails if post-explosion hydrodynamics prefers the recycling of some nucleosynthetic products over others. Furthermore, to fully exploit the stellar-archaeological program of constraining the Pop III initial mass function from the observed Pop II abundances, considering these hydrodynamical transport effects is crucial. We discuss applications to the rich chemical structure of ultra-faint dwarf satellite galaxies, to be probed in unprecedented detail with upcoming spectroscopic surveys.

astro-ph.GA↗

arXiv2015

Confined Population III Enrichment and the Prospects for Prompt Second-Generation Star Formation

It is widely recognized that nucleosynthetic output of the first, Population III supernovae was a catalyst defining the character of subsequent stellar generations. Most of the work on the earliest enrichment was carried out assuming that the first stars were extremely massive and that the associated supernovae were unusually energetic, enough to completely unbind the baryons in the host cosmic minihalo and disperse the synthesized metals into the intergalactic medium. Recent work, however, suggests that the first stars may in fact have been somewhat less massive, with a characteristic mass scale of a few tens of solar masses. We present a cosmological simulation following the transport of the metals synthesized in a Population III supernova assuming that it had an energy of 1e51 ergs, compatible with standard Type II supernovae. A young supernova remnant is inserted in the first star's relic HII region in the free expansion phase and is followed for 40 Myr employing adaptive mesh refinement and Lagrangian tracer particle techniques. The supernova remnant remains partially trapped within the minihalo and the thin snowplow shell develops pronounced instability and fingering. Roughly half of the ejecta turn around and fall back toward the center of the halo, with 1% of the ejecta reaching the center in 30 kyr and 10% in 10 Myr. The average metallicity of the combined returning ejecta and the pristine filaments feeding into the halo center from the cosmic web is 0.001 - 0.01 Z_sun, but the two remain unmixed until accreting onto the central hydrostatic core that is unresolved at the end of the simulation. We conclude that if Population III stars had less extreme masses, they promptly enriched the host minihalos with metals and triggered Population II star formation.

astro-ph.CO↗

arXiv2015

Metal Transport and Chemical Heterogeneity in Early Star Forming Systems

To constrain the properties of the first stars with the chemical abundance patterns observed in metal-poor stars, one must identify any non-trivial effects that the hydrodynamics of metal dispersal can imprint on the abundances. We use realistic cosmological hydrodynamic simulations to quantify the distribution of metals resulting from one Population III supernova and from a small number of such supernovae exploding in close succession. Overall, supernova ejecta are highly inhomogeneously dispersed throughout the simulations. When the supernova bubbles collapse, quasi-virialized metal-enriched clouds, fed by fallback from the bubbles and by streaming of metal-free gas from the cosmic web, grow in the centers of the dark matter halos. Partial turbulent homogenization on scales resolved in the simulation is observed only in the densest clouds where the vortical time scales are short enough to ensure true homogenization on subgrid scales. However, the abundances in the clouds differ from the gross yields of the supernovae. Continuing the simulations until the cloud have gone into gravitational collapse, we predict that the abundances in second-generation stars will be deficient in the innermost mass shells of the supernova (if only one has exploded) or in the ejecta of the latest supernovae (when multiple have exploded). This indicates that hydrodynamics gives rise to biases complicating the identification of nucleosynthetic sources in the chemical abundance spaces of the surviving stars.

astro-ph.GA↗

arXiv2015

Star formation in the first galaxies - III. Formation, evolution, and characteristics of the first stellar cluster

We simulate the formation of a low metallicity (0.01 Zsun) stellar cluster in a dwarf galaxy at redshift z~14. Beginning with cosmological initial conditions, the simulation utilizes adaptive mesh refinement and sink particles to follow the collapse and evolution of gas past the opacity limit for fragmentation, thus resolving the formation of individual protostellar cores. A time- and location-dependent protostellar radiation field, which heats the gas by absorption on dust, is computed by integration of protostellar evolutionary tracks with the MESA code. The simulation also includes a robust non-equilibrium chemical network that self-consistently treats gas thermodynamics and dust-gas coupling. The system is evolved for 18 kyr after the first protostellar source has formed. In this time span, 30 sink particles representing protostellar cores form with a total mass of 81 Msun. Their masses range from ~0.1 Msun to 14.4 Msun with a median mass ~0.5-1 Msun. Massive protostars grow by competitive accretion while lower-mass protostars are stunted in growth by close encounters and many-body ejections. In the regime explored here, the characteristic mass scale is determined by the temperature floor set by the cosmic microwave background and by the onset of efficient dust-gas coupling. It seems unlikely that host galaxies of the first bursts of metal-enriched star formation will be detectable with the James Webb Space Telescope or other next-generation infrared observatories. Instead, the most promising access route to the dawn of cosmic star formation may lie in the scrutiny of metal-poor, ancient stellar populations in the Galactic neighborhood. The observable targets that correspond to the system simulated here are ultra-faint dwarf satellite galaxies such as Bootes II, Segue I and II, and Willman I.

astro-ph.GA↗

arXiv2014

Recovery from population III supernova explosions and the onset of second generation star formation

We use cosmological simulations to assess how the explosion of the first stars in supernovae (SNe) influences early cosmic history. Specifically, we investigate the impact by SNe on the host systems for Population~III (Pop~III) star formation and explore its dependence on halo environment and Pop~III progenitor mass. We then trace the evolution of the enriched gas until conditions are met to trigger second-generation star formation. To this extent, we quantify the recovery timescale, which measures the time delay between a Pop~III SN explosion and the appearance of cold, dense gas, out of which second-generation stars can form. We find that this timescale is highly sensitive to the Pop~III progenitor mass, and less so to the halo environment. For more massive progenitors, including those exploding in pair instability SNe, second-generation star formation is delayed significantly, for up to a Hubble time. The dependence of the recovery time on the mass of the SN progenitor is mainly due to the ionizing impact of the progenitor star. Photoionization heating increases the gas pressure and initiates a hydrodynamical response that reduces the central gas density, an effect that is stronger in more massive. The gas around lower mass Pop~III stars remains denser and hence the SN remnants cool more rapidly, facilitating the subsequent re-condensation of the gas and formation of a second generation of stars. In most cases, the second-generation stars are already metal-enriched to ~2-5 X 10^{-4}\zsun, thus belonging to Population~II. The recovery timescale is a key quantity governing the nature of the first galaxies, able to host low-mass, long-lived stellar systems. These in turn are the target of future deep-field campaigns with the James Webb Space Telescope.

astro-ph.GA↗

arXiv2014

Radiative Feedback from high mass X-ray binaries on the formation of the first galaxies and early reionization

Recent work suggests that the first generation of stars, the so-called Population III (Pop III), could have formed primarily in binaries or as members of small multiple systems. Here we investigate the impact of X-ray feedback from High-Mass X-ray Binaries (HMXBs) left behind in stellar binary systems after the primary forms a black hole (BH), accreting gas at a high rate from the companion, a process that is thought to be favored at the low metallicities characteristic of high-redshift gas. Thanks to their large mean free path, X-rays are capable of preionizing and preheating the gas in the intergalactic medium (IGM) and in haloes long before the reionization of the Universe is complete, and thus could have strongly affected the formation of subsequent generations of stars as well as reionization. We have carried out zoomed hydrodynamical cosmological simulations of minihaloes, accounting for the formation of Pop III stars and their collapse into BHs and HMXBs, and the associated radiation-hydrodynamic feedback from UV and X-ray photons. We find no strong net feedback from HMXBs on the simulated star formation history. On the other hand, the preheating of the IGM by HMXBs leads to a strong suppression of small-scale structures and significantly lowers the recombination rate in the IGM, thus yielding a net positive feedback on reionization. We further show that X-ray feedback from HMXBs can augment the ionizing feedback from the Pop III progenitor stars to suppress gas accretion onto the first BHs, limiting their growth into supermassive BHs. Finally, we show that X-ray ionization by HMXBs leaves distinct signatures in the properties of the high-redshift hydrogen that may be probed in upcoming observations of the redshifted 21cm spin-flip line.

astro-ph.CO↗