Development Of Potential Energy Surfaces Research Articles

How to train a neural network potential.

The introduction of modern Machine Learning Potentials (MLPs) has led to a paradigm change in the development of potential energy surfaces for atomistic simulations. By providing efficient access to energies and forces, they allow us to perform large-scale simulations of extended systems, which are not directly accessible by demanding first-principles methods. In these simulations, MLPs can reach the accuracy of electronic structure calculations, provided that they have been properly trained and validated using a suitable set of reference data. Due to their highly flexible functional form, the construction of MLPs has to be done with great care. In this Tutorial, we describe the necessary key steps for training reliable MLPs, from data generation via training to final validation. The procedure, which is illustrated for the example of a high-dimensional neural network potential, is general and applicable to many types of MLPs.

Diffusion Monte Carlo with fictitious masses finds holes in potential energy surfaces

Holes are commonly associated with high-dimensional, non-parametric representations of potential energy surfaces (PESs). These are regions of the PES with large negative values. Typically these occur at highly distorted geometries of (in reality) high energy. Finding holes in a systematic and efficient way is thus an important aspect in the development of PESs. Here we propose using diffusion Monte Carlo (DMC) with artificially reduced masses to automatically and robustly sample large regions of a fit, including those at very high energies. We demonstrate this for a recent PES of methane that was obtained using energies and gradients. The DMC software with the methane potential is available for download.

Ab Initio Potential Energy Surfaces and Quantum Dynamics for Polyatomic Bimolecular Reactions.

There has been great progress in the development of potential energy surfaces (PESs) and quantum dynamics calculations in the gas phase. The establishment of a fitting procedure for highly accurate PESs and new developments in quantum reactive scattering on reliable PESs allow accurate characterization of reaction dynamics beyond triatomic systems. This review will give the recent development in our group in constructing ab initio PESs based on neural networks and the time-dependent wave packet calculations for bimolecular reactions beyond three atoms. Bimolecular reactions of current interest to the community, namely, OH + H2, H + H2O, OH + CO, H + CH4, and Cl + CH4, are focused on. Quantum mechanical characterization of these reactions uncovers interesting dynamical phenomena with an unprecedented level of sophistication and has greatly advanced our understanding of polyatomic reaction dynamics.

GAFit: A general-purpose, user-friendly program for fitting potential energy surfaces

We have developed a software package based on a genetic algorithm that fits an analytic function to a given set of data points. The code, called GAFit, was also interfaced with the CHARMM and MOPAC programs in order to facilitate force field parameterizations and fittings of specific reaction parameters (SRP) for semiempirical Hamiltonians. The present tool may be applied to a wide range of fitting problems, though it has been especially designed to significantly reduce the hard work involved in the development of potential energy surfaces for complex systems. For this purpose, it has been equipped with several programs to help the user in the preparation of the input files. We showcase the application of the computational tool to several chemical-relevant problems: force-field parameterization, with emphasis on nonbonded energy terms or intermolecular potentials, derivation of SRP for semiempirical Hamiltonians, and fittings of generic analytical functions. Program summaryProgram title : GAFitProgram Files doi : http://dx.doi.org/10.17632/9gy6bjcwk5.1Licensing provisions : GNU General Public License 3 (GPL)Programming language : Fortran 90, C, Perl and JavaNature of problem : Potential energy surfaces (PESs) have a key role in reaction dynamics and molecular dynamics. Chemical dynamics simulations can be performed by using analytical PESs or by direct dynamics, that is, by computing energies and forces “on-the-fly” by molecular structure calculations. The development of analytical PESs is, quite often, a very complex and tedious task. There are several programs in the literature that may help to develop PESs but, as far as we know, none of them show, at the same time, great generality and flexibility.Solution method : GAFit was designed to help users develop analytical potential energy functions. We made special efforts to write a code that exhibits the three features mentioned above. For this reason, we have also interfaced GAFit with the CHARMM and MOPAC programs. The former is one of the popular packages for molecular dynamics and the latter is a semiempirical quantum chemistry program that may be conveniently used for direct dynamics simulations. Because, in general, the fitting of PESs involves a simultaneous optimization of many nonlinear parameters, we have selected a genetic algorithm as the driver for the parameterizations. As shown in the paper, GAFit may be applied to other fitting problems.