nap: A molecular dynamics package with parameter-optimization programs for classical and machine-learning potentials

Ryo Kobayashi
2021 Journal of Open Source Software  
The nap is a package for molecular dynamics (MD) simulation consisting of an MD program (pmd) that can perform large-scale simulation using a spatial-decomposition technique and two parameter-optimization programs: one for classical (CL) potentials (fp.py) and another for machine-learning (ML) potentials (fitpot). Since the numbers of parameters to be optimized are much different between CL and ML potentials, optimization approaches for them are also different; meta-heuristic global
more » ... ch algorithms for the CL potentials, in which the numbers of parameters are usually much less than one hundred, and gradient-based methods for the ML potentials. The parameters of CL potentials can be optimized to any target quantity that can be computed using the potentials since meta-heuristic methods do not require the derivatives of the quantity with respect to parameters. On the other hand, ML-potential parameters can be optimized to only energies, forces on atoms and stress components of reference systems, mainly because gradient-based methods require the derivatives of other quantities with respect to parameters, and the analytical derivatives and the coding of them are usually painful and sometimes impossible. Potentials can be used in combination with any other potential, such as pair and angular potentials, short-range and long-range potentials, CL and ML potentials. With using the nap package, users can perform MD simulation of solid-state materials with the choice of different levels of complexity (CL or ML) by creating interatomic potentials optimized to quantum-mechanical calculation data even if no potential is available. Statement of need MD simulation is widely used in many research fields such as materials science, chemistry, physics, etc., to study dynamics of atoms or molecules. In order to perform MD simulation of systems including large number of atoms, where quantum-mechanical calculations can not be used due to their computational cost, empirical interatomic potentials between species are required. And the results of MD simulation are strongly dependent on the property or accuracy of the potentials used in the simulation. Hence, there are a lot of CL potential models have been proposed such as Lennard-Jones (LJ) potential for van der Waals interaction, Coulombic potential for ionic interaction, Morse potential for covalent interactions (Morse, 1929), angular-dependent models for angles between covalent bonds, bond-order models for more complex systems, etc. Recently ML potentials have been also actively studied because they are usually more flexible and can reproduce reference data more precisely than CL potentials. Even though the potential is flexible or suitable to problems considered, the parameters in the Kobayashi, R., (2021). nap: A molecular dynamics package with parameter-optimization programs for classical and machine-learning potentials.
doi:10.21105/joss.02768 fatcat:7wbxbr7cmjfe3mtd5vjezm4azu