Over the last few decades, computational chemists and drug developers have been dependent on force fields for large scale simulation due to their favorable computational efficiency. While force fields have revolutionized computational studies, they tend not to offer the accuracy and transferability that computationally expensive quantum mechanical (QM) methods provide. Highly efficient machine learning-based methods have recently been shown to accurately learn QM calculated potential energies, as well as other properties such as the molecular dipole moment and atomic partial charges. Our recent work on developing the general purpose ANI neural network potential (NNP) for organic molecules has shown that, with the correct molecular descriptor and dataset, a NNP can be trained to predict energies and forces for the entire class of organic molecules. ANI is also shown to be capable of extending its predictions to molecular systems that are much larger than those included in the training dataset. Local chemical environment atom-centered descriptors, such as those used by ANI, perform very well at describing local atomic interactions when combined with neural networks; however, these models have traditionally ignored important long-range physics. In this work we introduce an all-in-one model that is trained to predict the molecular dipole moment and potential energy combined with long-range physics using an inferred charge model. We present results showing the accuracy and efficiency of the potential on atomistic systems including bulk water and organic/biologically relevant molecules. The subtle but important inclusion of long-range physics to general purpose ML potentials provide a force field efficient and QM accurate addition to computational chemist’s toolbelt, bringing the community a step closer to ending the age old tug of war between accuracy, transferability, and computational efficiency.

All-in-one general purpose ML potential with long-range physics

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.