(622f) Systematic Investigation of Training Protocols for Machine Learning Derived Interatomic Potentials

Molecular simulations have proved to be an important tool in understanding molecules and molecular interactions and have helped accelerate research in material discovery, catalysis design and biomedicine. Molecular dynamics (MD) simulations help understanding the conformational changes of molecules from the atomic configurations and obtaining potential energy surface (PES) for the systems. However, there is always a tradeoff between the accuracy and computational cost. While ab initio molecular dynamics (AIMD) offer improvements in the accuracy of results and preclude the need to pre-define molecular topology, poor scalability of electronic structure methods severely limits applicable system sizes. On the other hand, classical molecular dynamics methods significantly improve scalability, but are limited to low accuracy and other well-known limitations. To bridge this gap, there have been developments in using machine learning (ML) methods to develop an interatomic potential for molecules¹. Typical strategies seek to create ML potentials trained on AIMD data, which offer greater accuracy² while also being able to simulate larger, more relevant systems. However, there is a bottleneck in developing ML potentials that surrounds the breadth of required training data³ needed to create sufficient reconstruction accuracy of energies and forces. To overcome this challenge, we have explored using enhanced sampling methods to speed the generation of training data and improve the reconstruction accuracy of the potentials while also limiting the total amount of training data needed. We envision the applications of ML potentials on ion interactions with solvents involving long range interactions and on different ranges of concentration. We also illustrate the generalizability of ML potentials for liquids and explore the extent to which they can be used for systems of varied size and different thermodynamic ranges.

References:

1. Behler J, Parrinello M. Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces. Phys Rev Lett. 2007;98(14):146401. doi:10.1103/PhysRevLett.98.146401
2. Unke OT, Chmiela S, Sauceda HE, et al. Machine Learning Force Fields. Chem Rev. Published online March 11, 2021. doi:10.1021/acs.chemrev.0c01111
3. Vassilev-Galindo V, Fonseca G, Poltavsky I, Tkatchenko A. Challenges for machine learning force fields in reproducing potential energy surfaces of flexible molecules. The Journal of Chemical Physics. 2021;154(9):094119. doi:10.1063/5.0038516