(659f) Development of Physics-Informed Neural Network Potentials for Molecular Simulations

Wang, S. H., Virginia Tech
Achenie, L., Virginia Tech
Xin, H., Virginia Polytechnic Institute and State University
In molecular dynamics (MD) simulations, prediction of forces acting on atoms are dependent on force fields. In the past decades, numerous physically meaningful force field equations were proposed, such as the Lennard-Jones potential [1], CHARMM force field [2] and bond order potential (BOP) [3]. However, these models which have a few adjustable parameters have been limited in accuracy. In recent years machine learning has gained a lot of attention. For example the artificial neural network (NN) has been employed to predict potential energies and interatomic forces with high accuracy through interpolation among energies from the density functional theory (DFT) database [4]. However, most NN potentials lack enough physics in their mathematical construction leading to bad extrapolation ability and poor transferability to unseen systems. This paper combines NN potentials and classical force fields to improve transferability of NN potentials and higher accuracy than traditional force fields. The pure gold system is being studied with the physically-informed neural network potential (PINN) which has been demonstrated to give accurate results [5]. The algorithm is currently being extended to relevant catalytic systems, including water/gold interfaces.


[1] Jones, J. E. (1924). On the determination of molecular fields.—II. From the equation of state of a gas. Proceedings of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character, 106(738), 463-477.

[2] Brooks, B. R., Bruccoleri, R. E., Olafson, B. D., States, D. J., Swaminathan, S. A., & Karplus, M. (1983). CHARMM: a program for macromolecular energy, minimization, and dynamics calculations. Journal of computational chemistry, 4(2), 187-217.

[3] Tersoff, J. (1988). New empirical approach for the structure and energy of covalent systems. Physical Review B, 37(12), 6991.

[4] Behler, J., & Parrinello, M. (2007). Generalized neural-network representation of high-dimensional potential-energy surfaces. Physical review letters, 98(14), 146401.

[5] Pun, G. P., Batra, R., Ramprasad, R., & Mishin, Y. (2018). Physically-informed artificial neural networks for atomistic modeling of materials. arXiv preprint arXiv:1808.01696.