Deep learning has become a useful instrument in the scientist’s toolbox for tackling problems in computational chemistry. Neural Network Potentials (NNPs), such as the family of ANI potentials, rival the accuracy of correlated quantum mechanical (QM) methods such as CCSD(T) while requiring significantly lower computational cost and scaling. One of the challenges of these networks is an accurate description of long-range interactions. The Effective Fragment Potential (EFP) is a QM based method for modeling non-covalent interactions. EFP decomposes the non-covalent interaction energy into electrostatic, polarization, dispersion, exchange-repulsion and charge transfer terms by using perturbative expansions. EFP is also used as a polarizable force field in QM/MM models that have been successfully applied for describing environment effects on electronic structure and excited states. Here we describe an extension of ANI to a general neural network (NN) capable of modeling extended molecular systems. In this NN, electrostatic potentials generated by EFP are used as additional parameters for describing long-range interactions of the system of interest with the surrounding environment. In this proposed model, the long-range interactions are trained using EFP and QM/EFP models, while the short-range covalent interactions are adopted from accurate ANI NNPs. The developed computer codes will contribute to open-source <i>LibEFP</i> and <i>TorchANI</i> software libraries.

Extending ANI neural network potentials to macroscopic systems using EFP

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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