Educado
An EU Doctoral Network

There is a strong discrepancy between the number of low-mass dark matter haloes predicted by the current cosmological model and the number of low-mass dwarf galaxies observed. The reason behind this difference is unclear and, at present, we do not know whether this 20-year-old problem reflects a major complication on our galaxy formation paradigm or whether it is connected to a limited and biased view of the dwarf galaxies in our vicinity. We will use ultra-deep Subaru HSC-SSC, LBT, and LSST imaging data to analyse the population of satellites surrounding a sample of 100 Milky Way-like galaxies – a sample large enough so Poissonian statistics predict small enough uncertainties on the results. We will create luminosity and mass functions of such galaxies using a statistically representative number of objects, and compare them with state-of-the-art cosmological simulations (including, but not limited to, CLUES and HESTIA, which are Local Group simulations). This will be a major step forward towards solving the ‘missing satellites’ problem and its consequences on the nature of dark matter, and will also allow us to study any preferential orientation of the satellites. DC01 will first create a sample of 100 Milky Way-like host galaxies with HSC-SSC, LBT, and (at a later stage) LSST imaging. We will find and characterize the satellite galaxies around the 100 hosts down to a radius of 300 kpc and limiting surface brightness of 29 mag/arcsec2, and then build the luminosity and mass functions of satellite galaxies, an important problem in galaxy formation. To compare such functions with satellites from the Milky Way and M31, detailed comparisons at different surface brightnesses will be made with existing simulations. This will lead towards a clear understanding of the best models, in terms of baryonic physics and feedback, to be input in such simulations. DC01 will be seconded for 3 months at RUG to work with Wilkinson on the application of mathematical morphology algorithms, and for two months at DIRAC in the USA to work with Ivezic on aligning the algorithms and analysis to LSST data on dwarf galaxies.

Deep-sky astronomy explores distant celestial phenomena to advance cosmological models that explain the structure and evolution of the Universe. In this context, modern telescopes, with enhanced sensitivity and imaging modalities, are revealing new types of objects, beyond regular-looking galaxies, such as faint streams and arcs, but their properties remain poorly understood. This research project aims to develop innovative object detection and characterization techniques, leveraging cutting-edge machine learning and structural approaches for multi-modal and multi-dimensional astronomical imaging. These novel methods will provide interpretable and reliable semantic data on the wide variety of observable objects present in the latest deep-sky surveys. The DC will extend the earlier work on faint object detection with morphological filters and machine learning, in a more general framework, with better statistical models of the noise, more sophisticated instrumentation models and finally object models to significantly improve robustness. Results will include algorithms to detect objects in large multi-dimensional and multi-parametric data sets. The project includes two secondments: one at New York University (NYU) under the guidance of Professor Yann Lecun to work on supervised and self-supervised deep learning, and another at the Instituto de Astrofísica de Canarias (IAC) under the supervision of Johan Knapen to get acquainted with deep astronomical imaging.

DC03 will focus on understanding the role of environment in shaping the morphologies of low-mass galaxies. Working with DC02, we will apply state-of-the-art methods, based on a combination of mathematical morphology and deep learning, to characterize the morphologies of dwarf galaxies from massive new imaging surveys, in particular Euclid and LSST. Both surveys, because of their unprecedented sky coverage and depth, will enable DC03 to detect, for the first time, statistical samples of dwarf galaxies across different environments. While deep learning has been successfully applied to derive galaxy morphologies of massive galaxies, the dwarf regime is still poorly explored, mainly because of the poor statistics. We will build on the demonstrated expertise of our team and focus on techniques to reduce the size of the annotated datasets for training such as active and contrastive learning. We will also provide well-calibrated uncertainties based on Bayesian Neural Networks. In this project, the candidate will first produce a detailed morphological description of dwarf galaxies, including types of components, such as bulges, bars, and nuclear clusters, but also indicators of (tidal) interactions: faint asymmetries in the outer parts, or underlying spirals. The morphological properties in different environments will be compared, and deep learning-based methods developed for morphological characterization of dwarfs. Secondments of this project include one at RUG to work with Bunte on algorithms for deep learning, and a second at DIRAC to work with Connolly on applying ML algorithms to the analysis of deep LSST imaging.

The analysis of dwarf galaxies is expected to provide further insights into understanding structure formation in our universe. However, these small, low-mass Dwarf Galaxies, and remnants of their existence, are typically only measured in form of weak signatures and embedded in very complex heterogeneous observation and simulation data. The work of this project builds on earlier algorithms that we have applied to the cosmic web, GAIA data and simulations (Canducci et al. 2022). The student will extend and develop evolutionary computation techniques and game theory for the automatic detection of astronomical structures in spectra and images, incorporating features of weak signatures mined from simulations of DC07. This project is in close collaboration with DC11 in Birmingham, that develops probabilistic models of the detected structures. The student will have a first secondment in Gent at Ugent (supervisor De Rijcke) to learn about cosmological particle simulations of Milky Way analogues and to work with DC07, and an industrial internship of at Vicomtech (Spain) to work with Quartulli on modelling of structural features in heterogeneous multi-temporal Earth Observation data cubes for the detection and characterization of dynamic phenomena.

Ongoing and upcoming large photometric and spectroscopic stellar surveys are allowing astronomers to create a huge multidimensional map (position, space velocity and chemical composition) for ~10 million stars in the Halo, Disk(s) and Bulge of our Milky Way. The study of this unprecedented sample will allow to disentangle the series of events (and related physical processes) that led to the formation of the Galactic components. This is made possible by the fact that stars carry information about their past in their present-time chemistry and dynamical properties, so that by deconvolving the Galactic stellar populations through the lens of chemo-dynamics we can infer the characteristics of the early Milky Way. The objective of DC05 is to take advantage of this extraordinary dataset to focus on the study of the Halo, investigate the nature of the many substructures that have been identified in recent years and trace back groups of stars that formed together but since dispersed, ultimately distinguishing the accreted portion of the halo from that formed “in-situ”. The work will be done in close contact with DC06 and DC07. DC05 will develop and apply noise-reducing clustering and dimensionality-reduction algorithms that will allow to take advantage of the Halo dataset resulting from the combination of public data (Gaia, Gaia-ESO, APOGEE, GALAH) and proprietary data from the WEAVE HR Survey. The dataset for the inner Halo amounts to ~50K objects for which accurate full spatial phase information (3D position and velocities— x,y,z,Vx,Vy,Vz) is complemented by a thorough characterization of their composition (~15 elements) for a total of ~20-25 dimensions. For the outer Halo, the sample will be larger, ~200K, but with more limited dimensionality (<10). Detection in this multi-phase space of groups of stars born together and since dispersed (including completely or partially disrupted galaxies, stellar clusters, and stars lost from still existing clusters) will lead to a much better understanding of the formation of the inner Halo. Secondments are planned at RUG to work with Trager on clustering and dimensionality reduction in high dimensional data, and at Pervasive to work with Mor on the use of AI to attack complex and huge high-dimensional data sets.

DC06 will model recently discovered phase-space structures in the thin/thick disk Milky Way stellar system and analysed them as being produced both by pure internal dynamical excitations (bar and spirals) and by dynamical perturbations induced by the passage of dwarf galaxies such as the one of Sagittarius. We aim to unveil a top-down chemodynamical scenario for the formation and evolution of such structures combining both the unprecedented multi-dimensional Big Data provided by ongoing surveys (GAIA-DR3, WEAVE, 4MOST) and the novel and realistic MW analogues simulations developed in the project of DC07. The new codes inspired on the algorithms for deep learning (DL) and AI techniques for multi-modal data sources developed in the project of DC04 will be adapted and readjusted to unveil the correlations between the intricate dynamical agents acting on the system and the observed complex structures in the n-dimensional space of the observables (position, velocities, chemical abundances, and ages). Novel simulations boosting the required spatial and temporal resolution will allow DC06 to create a new dynamical framework. This will be used to develop a unique scenario of the young and perturbed MW disk highlighted by the Phase Spirals and the classical large-scale mechanisms of migration, heating and/or excitation of vertical breathing on each of the stellar components of the Galactic disk. Secondment are planned at RUG to work with Bunte on algorithms for Deep Learning applied to Galactic disk Gaia/WEAVE data, and at AVS to work with Salata on massive data processing technology in astronomy and Earth science.

DC07 will use a moving-mesh N-body/hydrodynamics code to perform cosmological zoom-simulations of Milky Way analogues and will compare the properties of the simulated Milky Way and its dwarf satellites with observations provided by the Gaia and WEAVE surveys. The candidate will go beyond the state-of-the-art in this field by using a hydrodynamics implementation well-suited for capturing the effects of ram-pressure stripping and by employing the novel SWIFT task-based parallelization paradigm. The code will be equipped with the sub-grid physics developed by the UGent dwarf galaxy group. The goal is to study in detail how the Milky Way and its dwarf galaxies affect each other’s evolution and how this galactic co-evolution is reflected by the observations. This will be done combining the results of DC05 and DC06 with the simulations of DC07. The candidate will first produce a catalogue of simulations of Milky Way analogues and their dwarf galaxy satellites. The DC will then perform a detailed study of the orbital distribution and chemical composition of the stars in the Milky Way disc and halo and compare these with data from the Gaia and WEAVE surveys. In close collaboration with DC04, we will develop the numerical tools to identify structures (spiral structures, streams, grooves, etc.) and other markers that link simulations with observations and constrain detection of weak signatures in the observational data. Secondments are at RUG to work with Bunte to get acquainted with structure capturing algorithms, and at ADCIS (with Gratin), a company specialized in computer vision, to get acquainted with methods for acquiring, processing, and analysing digital images, both from an academic and commercial viewpoint.

The aim of this project is to study the relation between the spectral and spatial information of nearby galaxies, with the ultimate aim to derive a considerable amount of physical information, which one usually obtains from spectroscopy, from much more easily obtainable and cheaper imaging. While IFU-spectroscopy is expensive to obtain, in terms of availability, observing time, and field of view, deep imaging, in contrast, allows observing fainter sources but discards the majority of the spectral information by integrating the light in dedicated photometric filters. As an example, younger galaxies show much more substructure than older ones. The PhD student will work deep IFU and deep imaging data. Several methods, including population synthesis and surface brightness fluctuations, will be used to obtain the information contained in the data. The student will work with student DC10, to develop a machine learning technique to relate spectral information with imaging. The student will have a 3-month secondment in Heidelberg at HITS (supervisor Polsterer) to learn about machine learning techniques applied to astronomical data, and an industrial internship of 3 months at Tilt (Groningen) with Valentijn on mining very large datasets to trace unauthentic behavior on social media.

Different types of multispectral imaging data are becoming common in many different disciplines including astronomy. Astronomical data, however, present specific issues such as high levels of noise and instrumental signatures which need to be addressed in a detailed way in order to cope with the data streams produced by a new generation of integral-field and multi-object spectroscopy. DC09 will explore novel ways to extract and parametrise signals in noisy astronomical data-cubes using machine learning (including Deep Learning) based methods, leveraging on the know-how and tools already existing in other fields (such as medical imaging, geophysics, etc.) and on recent attempts performed on ALMA data. DC09 will fine-tune and apply these methods to integral-field and multi-object spectroscopic data and catalogues from ALMA, WEAVE and MUSE, in collaboration with DC08 and DC10. By working in an environment which is already strongly involved in the field of XAI (Explainable Artificial Intelligence) the DC will pay special attention on how to render the results reliable and understandable for the astronomical community at large. DC09 will be seconded for three months to the IAC to work with Knapen on WEAVE datacubes of nearby galaxies, and for two months at CIRA (with Zazzaro) to generalize detection and validation in space research projects.

In order to understand evolutionary processes of galaxies, we have to sample from a rich spatial, spectral and temporal space of observations. Integral Field Units and Multi-Object Spectrographs allow us to jointly measure the spatial and spectral space. Unfortunately, IFUs are very constrained wrt. e.g., availability, observation time, source brightness. In contrast, deep imaging allows us to observe fainter sources and therefore in greater numbers. However deep imaging discards the majority of the spectral and temporal information by integrating over time with dedicated photometric filters. Our goal, together with DC08, is to reconstruct the information of the original space that is absent in the deep imaging. As this is an ill posed inverse problem, instead of a unique solution it has a set of possible reconstructions. This calls for a probabilistic treatment that captures the reconstruction uncertainty as a distribution, guided by the method of DC04. Thereby we obtain a comprehensive characterization of the observed galaxies, which enables us to not only perform morphological analysis but to additionally estimate physical parameters. While classical image segmentation and morphological classification approaches operate directly on images, the proposed approach is working in the latent space that underlies the observed images. This allows a physically motivated segmentation which incorporates domain knowledge. We will develop a model, based on data of the WEAVE and SAMI IFU surveys of dwarf galaxies, that learns how to map from multi-band images to spectral-spatial data-cubes. The model will then be applied to a sample of galaxies and the reconstructed spectral information will allow a detailed analysis of galaxy evolution. DC10 will spend significant time at RUG to work with Wilkinson on applying modern mathematical morphology approaches to astronomical integral field and deep imaging data, and at Spheer.AI (supervised by De Jong) to use AI pattern searches in Earth observation data.

The analysis of dwarf galaxies is expected to provide further insights into understanding structure formation in our universe. However, these small, low-mass Dwarf Galaxies, and remnants of their existence, are typically only measured in form of weak signatures and embedded in very complex heterogeneous observation and simulation data. The student will extend and develop probabilistic models of cosmological structures that are able to constrain the search space to the largest possible extend, by learning from detailed simulations with known ground truth. Simulations can provide rich astrophysical information at every point in time, revealing structures in the simulation data that can be translated into features of weak signatures to be mined in the complex and noisy observational data. The work of this project builds on earlier algorithms that we have applied to the cosmic web, GAIA data and simulations (Canducci et al. 2022). The student will further enhance the developed methods in the framework of learning with privileged information obtained from simulations and expert domain knowledge (e.g., stellar evolution theory). Therefore the student will work with DC07 in Gent, who provides detailed simulations of Milky Way analogues and in close collaboration with DC04 in Groningen, that develops detection techniques with evolutionary computation. The resulting algorithms will be used throughout the consortium.