Developing Machine-Learned Potentials for Coarse-Grained Molecular Simulations: Challenges and Pitfalls

Machine learning (ML) is having increasing impact in the physical sciences, engineering, and technology, addressing research problems that range from molecular reaction mechanisms to high-throughput screening of functional materials. One strategy for representing molecules mathematically is through the use of graphs, whose nodes and edges correspond to atoms and bonds or interatomic distances, respectively. By performing multiple convolution operations on a graph, each node can influence other, increasingly distant, nodes. The use of graph neural networks has recently shown great promise in the development of improved atomistic force fields, trained on quantum mechanical calculations [1]. On the other hand, the implementation of ML for the generation of coarse grained (CG) mapping schemes [2], [3], and CG force fields required for developing hierarchical multiscale modelling schemes [4] on the basis of atomistic simulations is a less explored topic, and their application to the study of complex bulk systems is still rare [2],[5].