ANAKIN-ME and Electrostatics

Kate Kelso Huddleston, Adrian E. Roitberg, Ramon Miranda-Quintana

University of Florida

ANAKIN-ME (Accurate NeurAl networK engINe for Molecular Energies, or ANI) is a transferable machine learning (ML) potential for organic molecules, successful at reproducing the chemical accuracy of the reference QM level of theory used in training. The ANI methodology supplies accurate predictions of QM reference calculations while balancing computational cost comparable to classical force fields, providing chemical insight and understanding into the energetics and dynamics of molecular systems for the purpose of studying materials and drug design.

In order, to extend the ANI methodology, we are currently developing a neural network potential (NNP) that will accurately predict electrostatics. Currently, we have successfully trained a model to accurately predict Minimal Basis Iterative Stockholder (MBIS) charges. The next step in this work is to expand the method to predicting not only atomic charges, but electronegativity and chemical hardness as well.

We are also working to develop a NNP that will accurately predict the energy of a system, while also providing atomic energy and atomic charge values. This partition of the molecule will be based on the total energy of the system, giving us our own energy-based partial charges. Because the atomic energy and charge values will directly correspond with one another, this will allow for the charge transfer models of conceptual DFT to be tested.