Exploring the substructure of nucleons and nuclei with machine learning

Perturbative quantum chromodynamics (QCD) ceases to be applicable at low interaction energies due to the rapid increase of the strong coupling. In that limit, the non-perturbative regime determines the properties of quarks and gluons (partons) in terms of parton distribution functions (PDFs) or nuclear PDFs, based on whether they are confined within nucleons or nuclei respectively. Related non-perturbative dynamics describe the hadronisation of partons into hadrons and are encoded by the fragmentation functions (FFs). This thesis focuses on the detailed study of PDFs in protons and nuclei as well as the charged pions FFs by means of a statistical framework based on machine learning algorithms. The key ingredients are the Monte Carlo method for error propagation as well as artificial neural networks that act as universal unbiased interpolators. The main topics addressed are the inference of proton PDFs with theoretical uncertainties and the impact on the gluon PDF from dijet cross sections; a global determination of nuclear PDFs exploiting the constraints from proton-lead collisions at the LHC and using for the first time NNLO calculations; a new determination of FFs from single-inclusive annihilation and semi-inclusive deep-inelastic scattering data; and a quantitative assessment of the impact of future colliders such as the High-Luminosity LHC and the Electron Ion Collider on the proton and nuclear PDFs.