Methodological advances in training bottom-up neural network coarse-grained force-fields

Bottom-up coarse-graining has recently been applied to parameterize neural network force-fields. However, progress in modeling small proteins has been slow. We here discuss two new methodological advances related to creating bottom-up coarse-grained neural network force-fields from reference atomistic data. These methods directly relate to improving and adapting Multiscale Coarse-Graining and generalized Yvon–Born–Green theory to the nonlinear data-hungry regime of neural network force-fields. We show that these approaches reduce the amount of data needed to accurately parameterize a coarse-grained model, sometimes with minimal additional computational cost.