Delli Veneri, Michele (2022) Resolution of Inverse Problems in Astrophysics through Deep Learning. [Tesi di dottorato]

Anteprima

Testo DelliVeneri_Michele_XXXV.pdf Download (29MB) | Anteprima

Abstract

Current and forthcoming Astronomical observatories are rapidly increasing the quantity, velocity and complexity of their data products pushing Astronomy in the Big Data regime. Extracting scientifically usable data from such instruments involves the resolution of ill-posed inverse problems traditionally solved with algorithms which cannot cope anymore with the rising complexity. In the last decade, Machine Learning has seen a deep rise in its use both within and outside Astronomy. In this Thesis, I have developed a set of Deep Learning (DL) based pipelines aimed at the resolution of two such problems: the Radio Interferometric Deconvolution, Source Detection and Characterisation problem for two different radio interferometers, the Atacama Large Millimeter/submillimeter Array (ALMA) and the Square Kilometer Array (SKA), and the TOLIMAN space telescope Astrometric signal detection problem. Given the novelty of the instruments and the need for controlled experiments for the development and comparison of solutions, all studies carried out in this Thesis use simulated data. SKA and TOLIMAN data were acquired through my participation in the SKA Data Challenge 2 and COIN TOLIMAN Focus meeting, while I developed a simulation framework able to generate the needed ALMA observations by levering parallel computing. The ALMA pipeline is composed of six DL models: a Convolutional Autoencoder (CAE) for source detection within the spatial domain of the integrated data cubes, a Recurrent Neural Network (RNN) for denoising and peak detection within the frequency domain, and four Residual Neural Networks (ResNets) for source characterisation. The detection performances of the pipeline were compared to those of other state-of-the-art methods within the field and significant improvements in performances and computational times are achieved. Source morphologies are detected with subpixel accuracies obtaining mean residual errors of 10^-3 pixels (0.1 mas) and 10^-1 mJy/beam on positions and flux estimations, respectively. Projection angles and flux densities are also recovered within 10% of the true values for 80% and 73% of all sources in the test set, respectively. A direct comparison with tCLEAN, the current image deconvolution method employed by CASA, the ALMA data reduction pipeline, is made on simplified mock data achieving a substantial improvement in reconstruction quality and speed. The SKA pipeline, which I developed to address the shortcomings of the baseline pipeline developed during the Challenge in collaboration with COIN, is based on a combination of a classical Compressed Sensing algorithm, my 3D implementation of the Multi Vision Model, with six DL models: A 3D CAE for source detection, a 3D ResNet classifier to detect and remove false detections, and four 3D ResNet regressors to predict sources morphological parameters. The performances of the debugged, re-trained and optimised baseline pipeline and the revised pipeline are compared with those of the other solutions to the challenge. The revised pipeline reaches the highest score with slight improvements over the challenge winners. The TOLIMAN pipeline is the only unsupervised pipeline developed in this Thesis and it is based on a CAE tasked with compressing the TOLIMAN image time series into a monodimensional latent space which is then analysed through a Lomb-Scargle periodogram in search of periodic components. The pipeline performances in detecting increasingly small and realistic Astrometric signals embedded within a series of simulated TOLIMAN observations of the Alpha Cen star system are compared to those of other sparsity-based state-of-the-art solutions within the field. The signals are simulated as time-dependent shifts in the positions of two overlapping point spread functions in the TOLIMAN images. Our pipeline is the only one which can reliably detect the signal with an amplitude of 10^-6 times the pixel size. The simulations contained only Poisson noise, in future works, all the more realistic sources of noise and systematic effects present in the real world satellites will be injected into the simulations.

Downloads

Actions (login required)

Modifica documento