Efficiently capturing long-range interactions for machine learned interatomic potentials

Molecular Dynamics (MD) simulations are a ubiquitous tool used for studying a variety of complex systems. The atomic forces required to run these simulations can be most accurately determined using quantum mechanical (QM) calculations. However, in practice the poor scaling of such QM methods preculdes their use for large scale soft matter simulations. The alternative emperical forcefields that exist can be evaluated significantly faster but suffer from relatively poor predictive accuracy and flexibility when compared with QM methods. Recently, the development of Machine Learned Interatomic Potentials (MLIPs) aims to balance the speed of classical forcefields with the accuracy and flexibility of QM methods. A number of effective MLIPs exist, however none combine all the necessary requirements for running accurate large scale simulations of soft matter systems. Specifcally, to efficiently simulate these systems the MLIPs must exhibit favourable scaling, parallelizability, generalizability, and efficiently capture complex long-range interactions which become significant at these larger length scales. We present such an MLIP that exihibits the advantages of existing strictly local MLIPs with the ability to efficiently capture long-range interactions using a novel information-passing scheme.