Some of the biggest open problems in modern cosmology are the nature of the cosmic dark sector, the discrepancy between the theoretically predicted versus the observed value of the cosmological constant, and the growing cosmological discordances and tensions between different observational probes. Notably, the Hubble constant, which describes how fast the Universe is expanding when measured locally, has an enormous statistical disagreement with that inferred from the early Cosmic Microwave Background data. These inconsistencies, in turn, necessitate the formulation of new physics beyond the standard cosmological model. Current and ongoing observations, together with upcoming surveys, will produce large volumes of data, whose accumulation and processing will require an upgradation and increase in the sophistication of our statistical tools before applying them to specific problems. Thus, we propose to build a deep learning architecture using advanced statistics in machine learning algorithms like neural networks to be integrated into cosmological community codes for emulated parameter inference. This will help us to select, in a model-independent way, generic features of some cosmological theories that satisfy all observations. Utilising the power of deep learning will be an ideal space to investigate new physics in the observational sector and discriminate between models that are degenerate in terms of current observational approaches, fostering the development of data-driven science as a valuable companion to the model-driven paradigm. The fellowship will contribute to the researcher's career development by acquiring advanced skills in machine learning approaches using Bayesian statistics and developing skills within the cosmological community through a series of events designed to disseminate his results to the broader public. The project will also serve to consolidate and extend the researcher's network of professional contacts within Europe and beyond.