NEMORA preprint reports long-range atomistic learning at hundreds of thousands of atoms

Researchers report in a preprint that NEMORA, a method for learning interactions between atoms over long distances, handles systems containing hundreds of thousands of atoms with computation time and memory that scale linearly with system size. The method learns how to pass information through an adaptive spatial hierarchy, extending the Fast Multipole Method with operators that capture interactions across distance scales. It can augment short-range models with or without symmetry constraints. On non-local benchmarks, the authors report force-error reductions of over an order of magnitude and energy-error reductions of up to three orders of magnitude relative to short-range backbones. Accuracy is better than or competitive with existing long-range extensions.