Towards Exact Molecular Dynamics Simulations with Machine-Learned Force Fields

The predictive power of molecular dynamics (MD) simulations hinges on the accuracy of the underlying interatomic potential, however ubiquitous classical force fields are typically challenged by quantum effects.
We demonstrate how to leverage symmetric gradient domain machine learning (sGDML) to reconstruct global molecular force fields from high-level ab initio calculations that faithfully represent the accuracy of the reference data. A key feature of our approach is its ability to exploit spatial and temporal physical symmetries in a fully data-driven way, which enables a detailed reconstruction even when the sampling is well below the Nyquist rate.
By doing so, the sGDML model can be parametrized from only a few hundred reference calculations and then allows converged MD simulations that provide insights into the dynamical behavior of molecules. We demonstrate how to reconstruct force fields for small molecules at the quantum-chemical CCSD(T) level of accuracy and outline how this process can be scaled to molecular solids using a hierarchical approach where intramolecular cohesive forces within the solid are reconstructed successively.