Photometric redshifts with machine learning, lights and shadows on a complex data science use case

The current role of data-driven science is constantly increasing its importance within Astrophysics, due to the huge amount of multi-wavelength data collected every day, characterized by complex and high-volume information requiring efficient and as much as possible automated exploration tools. Furthermore, to accomplish main and legacy science objectives of future or incoming large and deep survey projects, such as JWST, LSST and Euclid, a crucial role is played by an accurate estimation of photometric redshifts, whose knowledge would permit the detection and analysis of extended and peculiar sources by disentangling low-z from high-z sources and would contribute to solve the modern cosmological discrepancies. The recent photometric redshift data challenges, organized within several survey projects, like LSST and Euclid, pushed the exploitation of multi-wavelength and multi-dimensional data observed or ad hoc simulated to improve and optimize the photometric redshifts prediction and statistical characterization based on both SED template fitting and machine learning methodologies. But they also provided a new impetus in the investigation on hybrid and deep learning techniques, aimed at conjugating the positive peculiarities of different methodologies, thus optimizing the estimation accuracy and maximizing the photometric range coverage, particularly important in the high-z regime, where the spectroscopic ground truth is poorly available. In such a context we summarize what learned and proposed in more than a decade of research.