Enhanced force field construction for graphene monolayers via neural network-based fitting of density functional theory data

This study presents a novel neural network (NN) framework for developing force fields specific to graphene monolayers, utilizing data obtained from first-principles calculations. The authors analyze three primary force components, force magnitude and the cosines of two angles across different configurations of surrounding carbon atoms. Initially, the NN applied to the three nearest neighbors, achieving average absolute testing errors of 0.375 eV/Å, 0.092, and 0.085 for the respective components. Then, expanding the input variables to nine surrounding atoms, which significantly enhances the precision of the force field models, reducing the error in force magnitude to approximately 1%. This improvement represents a 33% to 59% increase in accuracy over the initial method. The results demonstrate the potential of NNs to generate highly accurate force fields for graphene.