The Planck collaboration has extensively used the six Planck HFI frequency maps to detect the Sunyaev-Zel'dovich (SZ) effect from clusters of galaxies by applying component separation to construct a full-sky map of the y parameter or matched multi-filters to detect galaxy clusters via the pressure associated with their hot gas. We present how different approaches based on machine learning (e.g. Artificial neural networks and deep learning) can be successfully used as alternatives to detect the SZ signal from the Planck HFI frequency maps. We show that not only Plank clusters of galaxies and new confirmed clusters are
recovered, but in addition, thousands of SZ sources are detected for which we have statistical indications of their cluster nature.

Strong gravitational lensing is a very powerful tool for probing dark matter properties, high-redshift galaxies, and for measuring cosmological parameters such as the Hubble constant H0. Forthcoming imaging surveys, and in particular the Rubin Observatory Legacy Survey of Space and Time (LSST), will revolutionize the field by increasing lens samples from about 1000 to nearly 100000 systems, which will open new avenues for statistical studies. Finding these new lens systems among the extensive data sets is however a challenging task. While this a classical binary classification problem, the rarity of strong lenses and the similar morphologies of spirals, ring or merging galaxies make it difficult to exclude interlopers. I will present our on-going developments of supervised machine learning algorithms for identifying strong lenses in deep multi-band imaging surveys. Firstly, I will describe our two-step procedure with a catalog-based neural network classification of source photometry, and a CNN trained on gri image cutouts, followed by a visual inspection. This method was applied to the 3 billion detections in the PanSTARRS 3pi survey and allowed us to identify 330 new high-quality candidates. This automated two-step classification allows for a systematic analysis of the entire sky and will be applicable to the ~4 mag deeper gri stacks from LSST. Secondly, I will present our on-going efforts to test network architectures in order to further increase the selection purity, minimize the visual inspection stage, and improve the classification efficiency.

Recent progresses in Machine Learning have unlocked new possibilities to tackle scientific problems by means of neural networks, and already many applications have been developed both in astrophysics and cosmology. In this presentation, using a Generative Adversarial Network (GAN), an unsupervised learning model, we demonstrate the possibility to learn the distribution of dark matter of the cosmic web, using as input the results of large-scale -dark matter only- simulations such as Gadget. We provide a statistical analysis showing that the neural networks learn the underlying distribution. Finally, we show that, using the generator learned by the GAN it is possible to design an Auto-Encoder. The AE is then capable of both inferring the latent code of the GAN from a snapshot of dark energy density, therefore opening the way to new applications such as inferring missing values on a corrupted dataset or deducing from density distribution at lower z its possible evolution at higher redshift.

Small scale CMB data contain a lot of cosmological information hidden in the different components : primordial CMB, tSZ effect, kSZ effect, CIB. Standard analyses assume templates for non primordial CMB component and lose the cosmological signature of large scale structures contained in secondary anisotropies.
I will present a new analysis of SPT data at small scales where the tSZ spectrum is derived from machine learning algorithms,  trained on the halo model, and brings additional constraints.  I will show the cosmological and scaling relation parameter constraints obtained by combining SPT CMB data and Planck tSZ measurements.

I will present our on-going work aimed at building unsupervised methods to analyse the dynamics of galaxies, particularly the Milky Way and nearby systems for which we have rich astrometric and radial velocity data. One of these methods is the ACTIONFINDER, an algorithm that is able to find automatically the coordinate mapping from phase space to action-angles in quite general potentials, given a set of orbit segments. I will also sketch out our current work trying to generalise the ACTIONFINDER to work directly on stellar stream data, and our efforts to recover the mass distribution in self-gravitating systems.

With: Benoit Famaey, Giacomo Monari, Foivos Diakogiannis and Wassim Tenachi

We present first measurements of photometric redshifts for UNIONS data. UNIONS u-, r-, i-, and z-band images are complemented with CFHTLenS-u, and Pan-STARRS medium-deep catalogues. For training and validation we make use of the DEEP2+3 and SDSS/eBOSS spectroscopic data in the UNIONS footprint. We use various machine-learning techniques to learn the mapping between galaxy redshifts and colours, including random forests, support vector regression, boosted decision trees, neural networks, and dictionary learning. We compare those results to the template-fitting method LePhare. To improve results, we add galaxy morphology parameters such as surface brightness, size, elongation, or the Gini coefficient, measured on PSF-corrected r-band images using ShapePipe.

Weak gravitational lensing is one of the most promising tools of cosmology to constrain models and probe the evolution of dark-matter structures. Yet, the current analysis techniques are only able to exploit the 2-pt statistics of the lensing signal, ignoring a large fraction of the cosmological information contained in the non-Gaussian part of the signal. Exactly how much information is lost, and how it could be exploited is an open question.

In this work, we propose to measure the information gain from using higher-order (i.e. non-Gaussian) statistics in the analysis of weak gravitational lensing maps. To achieve this goal, we implement fast and accurate lensing N-body simulations based on the TensorFlow framework for automatic differentiation. By implementing gravitational lensing ray-tracing in this framework, we are able to simulate lensing lightcones to mimic surveys like the Euclid space mission or the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). These simulations being based on differentiable physics, we can take derivatives of the resulting gravitational lensing maps with respect to cosmological parameters, or any systematics included in the simulations.  Using these derivatives, we can measure the Fisher information content of various lensing summary statistics on cosmological parameters, and thus help maximize the scientific return of upcoming surveys. 

Recent years have seen the emergence of a number of very efficient use of deep learning technology for cosmological inference. Notably cosmological emulators show great promises to generate very quick mock catalogs either from cheap dark matter simulations or directly from random numbers. Investigations on these emulators also highlight that, provided the probed cosmology for the purpose of generating mock samples is not too far away from the one used in the training samples, they represent a fair alternative to a full cosmological simulation. I will review a few of those emulators that were developped in the French community (e.g. Ramanah et al. 2019, 2020; Peirani et al. 2013, 2021 in prep.) and beyond (He et al 2019). I will link those to the global effort of doing a full Bayesian analysis of large scale structures with different tracers and different epochs, the BORG framework (Jasche & Wandelt 2013) or the IMNN framework (Charnock et al. 2018).