ReaxFF-nn: A Reactive Machine Learning Potential in GULP and the Applications in the Thermal Conductivity Calculation of Carbon Nanostructures

ReaxFF -nn stands for Reactive Force Field ( ReaxFF ) with neural networks and is currently added to the General Utility Lattice Program (GULP) by modern FORTRAN programming. With GULP and ReaxFF -nn with parameters that are trained by our I-ReaxFF package, the thermal properties, crystal properties, energy minimization, etc., can be done with precision at the quasi-density functional theory (DFT) level. Compared to other Machine Learning Potentials (MLPs), we do not construct a thoroughly new machine learning potential, but just simply used a small neural network for the bond order and bond energy calculations, and the uncorrected bond orders are chosen as the input atomic feature vector. The advantage of To validate the model which we have coded in the GULP, we compared the forces of a random structure between the auto-differentiate package and our FORTRAN code, and the difference between them is about 10 -6 . Furthermore, we provided a systematical study of the thermal conductivity ( κ ) of graphene and carbon nanotubes (CNTs) through the phonon Boltzmann transport equation. The value of κ thermal conductivity of graphene is very close to the DFT calculations. Therefore, we declare that the potential we have trained for sp2 carbon can reach the quasi-DFT level. In this work, we report the thermal conductivity of CNTs calculated by ReaxFF-nn at quasi-DFT level are range from 107.032 to 310.019 W.m -1 .K -1 .