Covalent organic frameworks (COFs) are a class of advanced materials that can be precisely engineered for diverse applications, including catalysis, flexible electronics, and sensors. However, COFs synthesised experimentally often exhibit a variety of structural defects and grain boundaries, which affect their properties. Because of their large and complex structure, COFs pose a considerable challenge for traditional ab-initio methods. Machine learning interatomic potentials (MLIPs) can be used to significantly accelerate property calculations, while retaining near ab-initio accuracy. Our team have parametrised an MLIP using the MACE architecture and a dataset of non-equilibrium confirmations of 2D COFs. We assessed the transferability of the MACE model computing atomic forces and phonon dispersions of unseen COFs, and compared these results to ReaxFF and reference data by Density Functional Theory using VASP code. Using the parametrised model, we explore the effect of defects and grain boundaries on thermal and elastic properties of COFs.