Metal-organic frameworks (MOFs) are modular networks made up of metallic bricks and organic ligands. These materials possess unique and interesting properties that make them highly promising for a myriad of applications in industry, e.g., gas sorption, catalysis and nanosensing devices. In the last decade, tremendous computational efforts were invested in attempts to unravel the mechanisms governing macroscopic MOF behaviour. However, current interaction potentials are still limited to unrealistic, periodic structures that neglect spatial disorder at the mesoscale and leave a significant void between simulation and experiment. Machine learning potentials (MLPs), which are trained on quantum mechanical data, show a lot of potential to simulate systems at longer length and time scales. Yet, the accuracy of a MLP depends crucially on the accuracy of structures in its training dataset. Generating such datasets is a major bottleneck in the development of MLPs for complex materials. In this project, we propose a novel active-learning approach that takes advantage of the modular nature of MOFs and drastically lowers the computational demands for MLP training. This approach will be used to uncover the mechanical properties of MOFs. For the first time, we will investigate spatial disorder in MOFs on the mesoscale at quantum accuracy, and construct a universal MLP transferable across a wide array of MOFs, surpassing the state-of-the-art along a spatial and configurational dimension.