Abstract
<jats:p>Activation of metal–organic frameworks (MOFs), which involves the removal of residual solvent species, is a crucial process in converting a raw MOF structure into a computation-ready structure. The pore channels can be altered by residual solvent molecules, which in turn substantially affects the adsorption performance of MOFs. In addition, the risk that solvent removal may lead to framework collapse should not be overlooked. Herein, we developed an automated structural post-processing workflow comprising two complementary modules: a graph neural network (GNN)-based module for solvent removal and a random forest (RF)-based module for predicting MOF framework stability after desolvation. In the GTsR module, the developed GNN models for the datasets involving free- and all- (free and bonded) solvent removal achieved the area-under-the-curve (AUC) values of 0.9993 and 0.9970 on the test set, respectively. We further analyzed the effects of residual solvents on the geometric descriptors, chemical features, and adsorption performance of MOFs. In the complementary RFbased module, achieving a test set AUC of 0.8911 enables a more informed assessment of whether MOF framework is likely to remain stable after desolvation. Benchmarking against existing methods shows that the integrated GTsR–RF workflow is accurate and computationally efficient. It provides a practical route for the preparation of stabilityaware, computation-ready MOF structures.</jats:p>