Talks

------------------------------------------------------------------

Session 1: Extragalactic Astronomy

Ashley Spindler - Variational Inference or: How I Learned to Stop Supervising and Love the Bayes

• When it comes to extragalactic astronomy, machine learning methods have typically followed the same process: given some set of images that has labels from expert classification or citizen science, how can we best train a neural network to predict these labels? For example, what’s the best way to build and train a CNN to predict galaxy morphology based on the Galaxy Zoo classification scheme? But, what if your dataset isn’t labelled? What if you want to build a system that can adapt to new kinds of data? What if your labelling scheme is flawed in some way? In short, what if we can’t supervise our network?

In this talk, I will introduce “Variational Inference”, which allows for powerful generative and clustering methods on a wide variety of data, and demonstrate how it works with the galaxy zoo dataset. We start with a simple assumption, that our data has some underlying low dimensional manifold, and we can map that manifold to some statistical distribution. From here, I will show how a Variational Autoencoder can be built to find classes of galaxies, via unsupervised clustering, and generate brand new images of galaxies by sampling from the underlying distribution.

Michael Smith - Generative deep fields: arbitrarily sized, random synthetic astronomical images through deep learning

• Generative Adversarial Networks (GANs) are a class of artificial neural network that can produce realistic, but artificial, images that resemble those in a training set. In typical GAN architectures these images are small, but a variant known as Spatial--GANs (SGANs) can generate arbitrarily large images, provided training images exhibit some level of periodicity. Deep extragalactic imaging surveys meet this criteria due to the cosmological tenet of isotropy. Here we train an SGAN to generate images resembling the iconic Hubble Space Telescope eXtreme Deep Field (XDF). We show that the properties of ``galaxies'' in generated images have a high level of fidelity with galaxies in the real XDF in terms of abundance, morphology, magnitude distributions and colours. As a demonstration we have generated a 7.6-billion pixel `generative deep field' spanning 1.45 degrees. The technique can be generalised to any appropriate imaging training set, offering a new purely data-driven approach for producing realistic mock surveys and synthetic data at scale, in astrophysics and beyond.

Maria del Carmen Campos Varillas - XID+: The probabilistic de-blender for confusion dominated maps

• The Herschel Extragalactic Legacy Project (HELP, Oliver et al., in prep.) main goal is the creation of a multi-wavelength data set of galaxies at high redshift, focusing on data from 23 fields observed by Herschel Space Observatory, which is able to observe at wavelengths (far-infrared and sub-millimetre) where maximum dust emission from high redshift galaxies occurs. The downside is the confusion noise and the poor spatial resolution at this wavelengths. The availability of data with higher spatial resolution at shorter wavelengths enables the de-blending of Herschel maps, and the creation of statistical samples to study galaxy formation. The creation of 170 million sources data set selected at shorter wavelengths (Shirley et al., 2019) with homogenised photometry has been a really valuable prior information to obtain photometry on far-infrared confusion dominated maps. As part of HELP, XID+ (Hurley et al., 2017) is a prior-based source extraction tool to perform photometry at the positions of known sources in far-infrared. XID+ is based on a probabilistic Bayesian framework. It allows the inclusion of prior information to obtain the full posterior probability distribution on flux estimates. Making use of the extensive HELP data we can built an information-rich probabilistic hierarchical model that can be use to estimate statistical properties of galaxy populations directly from the data.

Ryan Jackson - Unveiling the hidden universe: studying low-surface-brightness galaxies using the New Horizon cosmological simulation

• Galaxy evolution studies have been dominated by objects that lie above the surface brightness (SB) limit of current wide surveys like the SDSS ( 23 mag/arcsec^2). However, hints of a much larger population of low-surface-brightness galaxies (LSBGs) have recently been discovered. State-of-the-art cosmological simulations, and new deep wide surveys, have begun to show that these galaxies might dominate the local number density, indicating that our current understanding of galaxy evolution is incomplete. We use New Horizon, a cosmological hydro-dynamical simulation, to quantify the origin of LSBGs. We show that the majority of galaxies occupy a fairly tight LSB locus in the SB vs stellar mass plane, which is invisible in past surveys. However, some galaxies scatter off this locus, and these are the ones that are visible in past datasets and on which our theoretical paradigm is predicated. Key to understanding galaxy evolution, therefore, is to understand both the formation of the LSB locus and the objects that lie off it. Doing so allows us to make testable predictions for the next generation of deep-wide surveys, such as LSST.

Matthew David Alexander Orkney - A Core in an Ultra-faint Dwarf Galaxy

• Cosmological simulations have grown in size from hundreds of particles in the 1960s to billions of particles today, leading to a similarly rapid growth in storage and data mining needs. As resolution increases, new physics are resolved and sub-grid physics recipes must be carefully changed. The EDGE (Engineering Dwarfs at Galaxy formation’s EDGE) simulation suite tackles this problem from the “bottom up”, modelling increasingly large galaxies at a fixed spatial resolution of ~12 light years. We analyse the formation of two ultra-faint EDGE galaxies, which spontaneously form their own star clusters. We show how the formation of a central dark matter core progresses inside these galaxies, and the implications this has on the survival of the star clusters. Then, we compare our results to Eridanus II, an ultra-faint dwarf and the faintest galaxy known to host its own star cluster. With the advent of huge astronomical surveys like LSST and Euclid, astronomers will be able to probe low-brightness galaxies out to the scale of megaparsecs. Our current cosmological model predicts hundreds to thousands of new Eridanus II type galaxies. Investigating the formation and evolution of such galaxies will be crucial in our understanding of the universe.

Bonny Barkus - Walk this Way: Tracing the Paths of Radio Jets

• Cross-identification of radio sources with optical and infrared catalogues is essential for determining host properties and distances, leading to intrinsic properties such as luminosity and size; but it is also far from straight forward. For simple, compact or isolated sources this can be done in an automated fashion. However, for extended sources or those which contain multiple components this becomes more complicated and has more often been achieved through human classification. As surveys become larger and sources more numerous this method becomes less efficient. The LOFAR Two metre Sky Survey (LoTSS) is the largest radio survey to date in terms of numbers of sources and data volume and is sensitive to both compact and extended emission, making it ideal for the study of radio sources. Using the first data release from LoTSS, and studying the morphology of the largest, brightest, more complex structures; the innovative idea of ridgelines, tracing the path of a jet, to link radio sources to their host galaxies has been applied. I will discuss the ongoing work and potential of this method for improving automated identification of the host galaxies of radio sources.

Maria Georganti - Is WZ Sge a magnetic propeller?

• Dwarf novae (DNe) are a class of accreting white dwarf binary systems where a cool dwarf (late-type star) overflows its Roche lobe and transfers matter into a white dwarf through an accretion disk. These binaries undergo numerous outbursts, during which their brightness increases by several magnitudes (3-4 mag) in a short period of time, ranging from days to years. These outbursts are thought to be caused by an instability in the accretion disk which triggers the ionization of the material on it, known as the “Disk Instability Model” . WZ Sge constitutes the prototype of a subclass of these systems since it appears impressive super-outbursts, 2 to 3 times brighter than the standard DNe, having an unusually long timescale of 33 yr. These eruptions are characterised by a primary burst, followed by a knee in their decline and subsequent mini-outbursts until the system returns, eventually, into quiescence. As these “dips” and “re-brightenings” in the lightcurves of DNe are not yet fully comprehended, we investigate the magnetic propeller mechanism as a plausible scenario in case of WZ Sge. Particularly, the appearance of these “dips” and consequently, the decrease of the luminosity can be interpreted as being due to the onset of the magnetic propeller and the inhibition of accretion until the point where the propeller stops and we observe a “re-brightening”. In this talk, I will present time-resolved, high-resolution UV spectroscopic observations taken with Hubble Space Telescope (HST/STIS) during the 2001 super-outburst of WZ Sge, which cover both the declining and re-brightening phase of this system. We test our hypothesis by constructing the RMS spectra and we search for: i) peculiar line characteristics so as to disentangle the stable and variable components of the binary as we expect the magnetic propeller to vary and the white dwarf to be stable with respect to time, ii) spectroscopic evidence of the spin period of the white dwarf (Pspin ≈ 28 s) in our observations. Finally, I will discuss the implications of these findings relating to the other known white dwarf propelling system, AE Aqr, which is described by an unusual UV spectrum and how this study contributes to the broader image we have for accreting binaries.

Jakub Orwat-Kapola - Deep learning for time series analysis in observational astronomy and smart city industry

• Next generation astronomical surveys and the rapidly developing Internet of Things will soon enter the era of Big Data. These technologies all use sensors capable of generating multidimensional time series data at unprecedented rates, as they monitor the evolution of complex systems over time. This data deluge necessitates the development of automated methods of outlier detection, in order to flag events for follow up analysis. Unusual and anomalous observations carry invaluable scientific insights and operational intelligence for urban analytics, while being hidden in Terabytes of data. My research interest focuses on the application of deep learning to the task of outlier detection in time series data in observational astronomy and smart city industry. Ongoing experiments with deep autoencoders show promising results for the approach of outlier detection using latent variables extracted from light curves in an unsupervised manner. Diverse modes of activity of black hole x-ray binary GRS1915+105 were analyzed in the latent space to visualize the evolution of its variability and to identify unusual observations. Future work will involve the analysis of further data sets, including records of the rates of electricity usage, which will test the interdisciplinary applicability of the proposed method.

Paula Soares - Probing Cosmology with Realistic Intensity Mapping Simulations

• With upcoming radio surveys such as the SKA, the technique of Intensity Mapping will allow us to trace how neutral hydrogen clusters across the sky as a direct probe of how dark matter clusters. Measuring the power of this clustering at different physical scales can help constrain cosmological parameters to unprecedented precision, break degeneracies between them, and allow us to test our current cosmological model. For this to be accomplished with upcoming surveys, we must carefully study and prepare for observational effects which will affect our measurements, and possibly bias any parameters we wish to estimate from data. Of particular importance is the effect of foregrounds, since when removed these can remove some of the underlying signal. We can create realistic simulations of what we will observe with Intensity Mapping where we know the fiducial cosmology, and test how observational effects affect the cosmological signal. We can also perform MCMC analyses using different models to check whether we recover the fiducial cosmological parameters before and after including observational effects. We find that foreground removal greatly biases the measurement of parameters for a simple cosmological model, and must be correctly accounted for in upcoming surveys.¬

------------------------------------------------------------------

Session 2: Inside the Milky Way

Simone Scaringi - Linking accretion across the scales with Kepler and TESS

• From planets to super-massive black holes, accretion (the accumulation of matter on a self-gravitating body through gravity) is the process by which most objects in the Universe grow in mass. Accretion requires angular momentum to be lost from the in-falling material, usually resulting in the formation of a so-called accretion disk. Although the importance of accretion disks have been recognized for many years, the detailed physics and dynamics are still poorly understood. Over the last decade we have been able to link the accretion physics of stellar-mass black holes with those of super-massive black holes, with over nine orders of magnitude difference in mass. However, we do not yet know if the physics of accretion can be extended to include other systems, such as accreting white dwarfs, neutron stars, and young-stellar objects. Although seemingly different observationally, I will show how all these different types of accreting systems have also revealed strikingly similar properties. Being just the "tip of the iceberg", the discoveries I will present suggest that a single unifying physical model might exist to explain how accretion disks behave throughout the Universe, irrespective of the mass, size, or type of the accreting object. If time allows I will also briefly touch on more recent efforts aimed at applying deep learning regression methods and attempt to map Kepler and TESS timeseries onto the Gaia colour-magnitude space.

Michele De Leo - Revealing the Tidal Scars of the Small Magellanic Cloud

• The ESA Gaia satellite Data Release 2, with it’s 1331909727 sources with full astrometric solution, is a perfect example of Big Data. I present in this talk the analysis of a Gaia DR2 subsample crossmatched with spectroscopic data of red giant branch stars of the Small Magellanic Cloud (SMC). The Large and the Small Magellanic Clouds (LMC/SMC), due to their proximity to us and turbulent interaction history, provide natural laboratories to understand how galaxies form, evolve and how tidal interactions shape these processes. I will show that the SMC centre of mass is clearly offset from the velocity centre of its associated HI gas, demonstrating that the latter is likely to be far from dynamic equilibrium. I will also present the first unequivocal confirmation that the SMC is currently undergoing tidal disruption by the LMC. I find evidence of tidally stripped stars, projected along the line-of-sight, well within 2 kpc of the photometric centre of the SMC. In order to understand how advanced is the state of tidal disruption in the SMC, I compare these findings to numerical models of the SMC/LMC system disrupting around the Milky Way. This work has been submitted as a paper to the MNRAS journal.

Benjamin Ross Law - Investigating our Galaxy with Runaway Stars

• In the era of big data astronomy, Gaia and it’s high precision proper motions and parallaxes can be used to simulate ‘Runaway Stars’ orbits through the galactic potential, to find their point and method (Dynamic and Binary Ejection mechanisms) of ejection, as well as potentially map spiral arms. These runaways can act as a proxy to hidden information for areas suffering from high extinction. Using theoretical atmospheres and published passbands of photometric instruments, we can create synthetic photometry for stars. The theoretical absolute magnitudes this produces can then be estimated for a given effective temperature, surface gravity, and reddening. This entails fitting routines that minimize the residual between the colour index of observations and models. Synthetic photometry will be used to select a potential sample of O, B, or A-type runaway stars from Gaia and 2MASS observations. Stellar evolution tracks can be applied to this sample to estimate their age and mass. With an estimated age, we can constrain the duration of the maximum flight time of the star and use this to constrain its origin.

Calum Morris - Using Time Domain Infra-Red Surveys to Identify Eruptive Variable YSO's

• The two widely separated epochs of UKIDSS K band photometry yielded a sample of several hundred variable young stellar object (YSO) candidates, many of these being in well studied SFRs. Here I present a follow up study of 31 of these objects, many of which are in the Cygnus X complex, using Subaru/IRCS and Gemini/NIFS 8m-class spectroscopy, In addition to NEOWISE time series photometry. The set features 1 confirmed FUOr type object, as well as 4 other FUOr-like stars, whilst we also present others that resemble EXors or other accreting YSOs. Some sources however display periodic Mid-IR Light curves resembling Long period variables (LBVs) or Dusty Mira-type AGB stars. This dichotomy in stellar classifications highlights how much we still need to learn of the infra-red variable sky.

Posters

Jaime Vargas-González - The Orion Radio All-Stars - VLA observations of clustering, proper motions, and variability in the ONC

• The study of nearby massive star forming regions in our Galaxy allows us to obtain a detailed picture of the kinematics and dynamics governing this kind of environments and thereby constrain the general perspective of star formation theories. The Orion Nebula Cluster (ONC) is the most prominent and ideal target for this purpose located at ~414pc from us. In this work we present a total of ~30 hr centimeter-wavelength observation of seven adjacent pointings over an area of ~15x15 arcmin centered on the ONC using the C-band (4 - 8 GHz) of the the Karl G. Jansky Very Large Array with its A-configuration. Additionally, we used previous VLA observations of the central part of the ONC to measure the proper motions of young stars with compact radio emission finding fast sources likely related to the BN/KL explosive event. From the most recent observations ~49% of the detections correspond to new sources not previously reported at centimetric wavelengths.