The molecular dynamics community is continuously working on the development of models and algorithms that are able to reproduce experimental properties. While ideally correlated methods such as coupled cluster would predict the closest value to experiments, its cost makes it unaffordable for large systems, and other less accurate approaches such as forcefields or density functional theory must be used instead. In this work we present a software infrastructure that allows us to generate and use MB-nrg potential energy functions (PEFs), which are based on the many-body expansion of the energy, and are an orthogonal approach that combines the accuracy of coupled cluster with the speed of a force field. The MB-nrg PEFs include classical dispersion and repulsion through a modified Buckingham potential, classical electrostatics and many-body polarization, supplemented by a set of permutationally invariant polynomials that correct for the deficiencies in the classical description due to the quantum effects arising from the overlap of the electronic density of the different monomers. The first part of the software, MB-Fit, allows the user to generate a training set for a given monomer, dimer, or trimer, along with the monomer properties such as charges, polarizabilities and C6 coefficients, which are derived purely form ab-initio methods, and the permutationally invariant polynomials. MB-Fit also provides a calculator that computes the n-body energies of the training set, performs the fits to reproduce those energies with a high accuracy, and adds the new surface into the MBX library. MBX is a C++ optimized and parallelized energy calculator for MB-nrg PEFs, and is interfaced with MD engines such as i-pi or LAMMPS. MBX can also be interfaced with python, fortran, and C. The combination of these two software packages can be an important step forward in providing the scientific community with the tools to easily generate PEFs for molecules of interest and predict experimental properties with a high accuracy.

Automatic generation and implementation of data-driven many-body models

ACS Spring 2022

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