Physics-based kinetic models are regarded as key tools for supporting the design and control of chemical processes and for understanding which degrees of freedom ultimately determine the observed behaviour of chemical systems. These models are formulated as sets of differential and algebraic equations where many state variables and parameters may be involved. Nonetheless, the translation of the available experimental evidence into an appropriate set of model equations is a time and resource intensive task that significantly relies on the presence of experienced scientists. Automated reactor platforms are increasingly being applied in research laboratories to generate large amounts of kinetic data with minimum human intervention. However, in most cases, these platforms do not implement software for the online identification of physics-based kinetic models. While automated reactor technologies have significantly improved the efficiency in the data collection process, the analysis of the data for modelling purposes still represents a tedious process that is mainly carried out a-posteriori by the scientist. This project focuses on how to systematically solve some relevant problems in kinetic modelling studies that would normally require the intervention of experienced modellers to be addressed. Specifically, the following challenges are considered: i) the selection of a robust model parametrisation to reduce the chance of numerical failures in the course of the model identification process; ii) the experimental design and parameter estimation problems in conditions of structural model uncertainty; iii) the improvement of approximated models embracing the available experimental evidence. The work presented in this Thesis paves the way towards fully automated kinetic modelling platforms through the development of intelligent algorithms for experimental design and model building under system uncertainty. The project aims at the definition of comprehensive and systematic modelling frameworks to make the modelling activity more efficient and less sensitive to human error and bias.