Galaxy clusters play an important role in modern cosmology and astrophysics. They act as cosmic laboratories where we can study galaxy formation and evolution, and improve our understanding of the nature of Dark Matter using dynamical and gravitational lensing methods. As powerful gravitational lenses, clusters can be used as natural cosmic telescopes thus extending our detection limit of faint sources and revealing the most distant galaxies. In this context, dedicated surveys with Hubble Space Telescope (HST) and ground-based extensive spectroscopic campaigns have provided data with extraordinary quality. The richness of these data sets, however, cannot be compared with the impressive data volume that upcoming surveys (such as Euclid, Vera Rubin Observatory or James Webb Space Telescope) will generate over the next years. The volume and the complexity of these new datasets can be efficiently dealt using machine learning and deep learning methods, which enable the exploration of hidden correlations within a multi-dimensional parameter space. In this thesis, we take advantage of this multidisciplinary tool to enable many scientific investigations of cluster internal structure and background source population. As a first application, we implemented deep learning architectures to select galaxy cluster members in galaxy clusters, in the redshift range 0.2 - 0.6, which is a critical first step for a variety of studies, such as galaxy evolution in dense environments, cluster mass estimates, strong lensing models. By using HST multi-band images alone, convolution neural networks (CNNs) were used to disentangle members from background and foreground sources, once they were trained with a large sample of spectroscopically confirmed sources (VLT VIMOS and MUSE observations), thus avoiding the complicated and time consuming photometric measurement process. We performed several experiments, finding that CNNs can classify members with a purity-completeness rate of ~90%, and showing stable results across the parameter space. As a second step, we focused on the identification of galaxy-galaxy strong lenses (GGSL) in galaxy clusters, which can be used to study the internal mass distribution of clusters, traced by the sub-halo population around cluster member galaxies, and can later be compared with cosmological simulations. In this work, we opted for a methodology that combines the need to simulate a large number of GGSL to train deep neural networks, while maintaining the imaging complexity of real observations. By exploiting high-precision cluster lens models available for 8 clusters (with redshift in 0.2 - 0.6), we used the estimated deflection angle maps to simulate thousands of realistic strong-lenses in real HST. We found that deep networks can detect a large fraction of real GGSLs, with a limited number of false negative events. We processed hundreds of members (spectroscopically confirmed or selected with the CNN), to test deep model generalization capabilities and to search for GGSL candidates. Finally, we implemented a 3D spectroscopy cross-correlation tool on the MUSE integral field spectrograph data to measure redshifts in an automated and computationally efficient fashion. The mining of spectroscopic information allows us to build datasets used to train neural networks, confirm cluster galaxy membership, measure the redshift of lens and source in lensing events. Optimized to be executed on graphic processors, this tool can process an entire MUSE dataset in a few tens of seconds, by cross-correlating 90000 spectra included in the data cube with a sample of spectral templates. Even though the tool is still under development our preliminary results appear rather promising and will soon be applied routinely on MUSE data. The methodologies developed in this thesis can be extended beyond the HST imaging data with a relatively modest effort and promise to have important applications with the upcoming next-generation facilities.

Gli ammassi di galassie hanno un ruolo importante nella cosmologia e nell'astrofisica moderne. Essi figurano come laboratori cosmici nei quali è possibile studiare la formazione e l'evoluzione delle galassie, e migliorare la nostra comprensione della materia oscura usando metodi basati sulla dinamica o su lenti gravitazionali. Come potenti lenti gravitazionali, gli ammassi agiscono da telescopi cosmici estendendo il limite di rilevamento di sorgenti deboli e rivelando galassie lontane. In questo contesto, survey dedicate con il Telescopio Spaziale Hubble (HST) ed estese campagne spettroscopiche hanno fornito dati di straordinaria qualità. Tuttavia, la ricchezza di questi dati non può essere paragonata all'impressionante volume che i futuri telescopi (come Euclid, Vera Rubin Observatory o James Webb Space Telescope) genereranno nei prossimi anni. Il volume e la complessità di questi nuovi dataset possono essere gestiti in modo efficiente con metodi di machine learning e deep learning, che consentono l'esplorazione di correlazioni nascoste all'interno di spazi multidimensionali. Come prima applicazione, abbiamo implementato architetture di deep learning per selezionare i membri di ammassi di galassie, con redshift in 0.2 - 0.6, un primo passo fondamentale per una varietà di studi, come l'evoluzione delle galassie in ambienti densi, stime di massa degli ammassi, modelli di strong lensing. Una volta addestrate con un ampio campione di sorgenti spettroscopicamente confermate (osservazioni VLT VIMOS e MUSE), le reti neurali convolutive (CNN) sono state utilizzate per separare i membri dalle sorgenti di background e foreground, utilizzando solo immagini multi-banda HST, evitando così il complicato e time-consuming processo di estrazione di misure fotometriche. Abbiamo eseguito diversi esperimenti, determinando che le CNN possono classificare i membri con un tasso di purezza-completezza ~90%, mostrando risultati stabili nello spazio dei parametri. Come secondo passo, ci siamo concentrati sull'identificazione dei galaxy-galaxy strong-lenses (GGSL) in ammassi, che possono essere utilizzati per studiare la distribuzione di massa degli ammassi, tracciare la popolazione di sub-aloni attorno ai membri. In questo lavoro, abbiamo optato per una metodologia che combina la necessità di simulare un gran numero di GGSL per addestrare reti neurali, mantenendo la complessità delle osservazioni reali. Abbiamo utilizzato le mappe di deflection angle stimate da modelli ad alta precisione di lensing dell’ammasso, disponibili per 8 cluster (con redshift in 0.2 - 0.6), per simulare migliaia di esempi realistici nelle immagini HST. Abbaimo determinato che le CNN possono rilevare un'ampia frazione di GGSL reali, con un numero limitato di falsi negativi. Abbiamo processato centinaia di membri (spettroscopicamente confermati o selezionati con la CNN), per testare la capacità di generalizzazione delle CNN e per cercare candidati GGSL. Infine, abbiamo implementato uno strumento di cross-correlazione 3D per i dati dello integral field spectrograph MUSE per misurare redshift in modo automatizzato e computazionalmente efficiente. L'estrazione di informazioni spettroscopiche ci consente di costruire dataset per addestrare reti neurali, confermare la membership di galassie, o misurare i redshift della lente e della sorgente in eventi di lensing. Ottimizzato per essere eseguito su processori grafici, questo strumento può elaborare un intero cubo MUSE in poche decine di secondi, cross-correlando 90000 spettri con un set di template. Anche se lo strumento è ancora in fase di sviluppo, i risultati preliminari sembrano piuttosto promettenti e saranno presto applicati di routine sui dati MUSE. Le metodologie sviluppate possono essere estese oltre i dati HST con uno sforzo relativamente modesto e promettono di avere importanti applicazioni per le imminenti survey di prossima generazione.

Deep Learning in Galaxy Clusters

ANGORA, Giuseppe
2022

Abstract

Galaxy clusters play an important role in modern cosmology and astrophysics. They act as cosmic laboratories where we can study galaxy formation and evolution, and improve our understanding of the nature of Dark Matter using dynamical and gravitational lensing methods. As powerful gravitational lenses, clusters can be used as natural cosmic telescopes thus extending our detection limit of faint sources and revealing the most distant galaxies. In this context, dedicated surveys with Hubble Space Telescope (HST) and ground-based extensive spectroscopic campaigns have provided data with extraordinary quality. The richness of these data sets, however, cannot be compared with the impressive data volume that upcoming surveys (such as Euclid, Vera Rubin Observatory or James Webb Space Telescope) will generate over the next years. The volume and the complexity of these new datasets can be efficiently dealt using machine learning and deep learning methods, which enable the exploration of hidden correlations within a multi-dimensional parameter space. In this thesis, we take advantage of this multidisciplinary tool to enable many scientific investigations of cluster internal structure and background source population. As a first application, we implemented deep learning architectures to select galaxy cluster members in galaxy clusters, in the redshift range 0.2 - 0.6, which is a critical first step for a variety of studies, such as galaxy evolution in dense environments, cluster mass estimates, strong lensing models. By using HST multi-band images alone, convolution neural networks (CNNs) were used to disentangle members from background and foreground sources, once they were trained with a large sample of spectroscopically confirmed sources (VLT VIMOS and MUSE observations), thus avoiding the complicated and time consuming photometric measurement process. We performed several experiments, finding that CNNs can classify members with a purity-completeness rate of ~90%, and showing stable results across the parameter space. As a second step, we focused on the identification of galaxy-galaxy strong lenses (GGSL) in galaxy clusters, which can be used to study the internal mass distribution of clusters, traced by the sub-halo population around cluster member galaxies, and can later be compared with cosmological simulations. In this work, we opted for a methodology that combines the need to simulate a large number of GGSL to train deep neural networks, while maintaining the imaging complexity of real observations. By exploiting high-precision cluster lens models available for 8 clusters (with redshift in 0.2 - 0.6), we used the estimated deflection angle maps to simulate thousands of realistic strong-lenses in real HST. We found that deep networks can detect a large fraction of real GGSLs, with a limited number of false negative events. We processed hundreds of members (spectroscopically confirmed or selected with the CNN), to test deep model generalization capabilities and to search for GGSL candidates. Finally, we implemented a 3D spectroscopy cross-correlation tool on the MUSE integral field spectrograph data to measure redshifts in an automated and computationally efficient fashion. The mining of spectroscopic information allows us to build datasets used to train neural networks, confirm cluster galaxy membership, measure the redshift of lens and source in lensing events. Optimized to be executed on graphic processors, this tool can process an entire MUSE dataset in a few tens of seconds, by cross-correlating 90000 spectra included in the data cube with a sample of spectral templates. Even though the tool is still under development our preliminary results appear rather promising and will soon be applied routinely on MUSE data. The methodologies developed in this thesis can be extended beyond the HST imaging data with a relatively modest effort and promise to have important applications with the upcoming next-generation facilities.
ROSATI, Piero
LUPPI, Eleonora
File in questo prodotto:
File Dimensione Formato  
PhD_Angora_2print.pdf

accesso aperto

Descrizione: PhD_Angora_2print.pdf
Tipologia: Tesi di dottorato
Licenza: DRM non definito
Dimensione 49.82 MB
Formato Adobe PDF
49.82 MB Adobe PDF Visualizza/Apri

I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11392/2481663
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact