Transferable and extensible machine learning derived atomic charges for modeling metal-organic frameworks

In most cases, calculations of properties of metal-organic frameworks (MOFs) based on classical force fields (FFs) are the most suitable in terms of the ratio between accuracy and computational cost, especially in efforts to screen a large number of structures. Such calculations require an initial partial charge assignment to describe the Coulomb contribution. In this study, we would like to present a machine-learning algorithm for MOF partial charge prediction and its verification on experimental data using the FF approach. Proposed ML method offers the accuracy of reference DFT calculations at a fraction of the computational cost with linear scalability.