Metal-organic frameworks (MOFs) are a recent class of scaffold-like materials consisting of inorganic building blocks connected through organic ligands. Their extraordinary physical properties make them ideally suited for applications such as gas separation or nano-actuation, and enormous computational research efforts have therefore been dedicated towards an understanding of the relation between the microscopic structure of the MOF and its macroscopic properties. Such structure-property relationships are traditionally obtained with computational models that treat MOFs as perfectly ordered and infinitely extending periodic structures, in sharp contrast with the spatially disordered microcrystals that are observed in actual experiments. To bridge this gap between theory and experiment, this proposal aims to develop new computational models that accurately capture both the inherent spatial disorder as well as the finite crystal size of real MOFs. In our bottom-up approach, we first employ a machine learning model to learn the atomistic interactions within the framework on a small scale (~1nm). These machine learning models are then used to parameterize accurate and transferable coarse-grained (CG) models on a larger scale (~10nm). Finally, these CG models are employed to computationally investigate stimuli-responsive behaviour in disordered finite crystals on an even larger scale, i.e. the experimental scale (~1µm).