A significant hurdle to studies of highly anharmonic molecules and clusters is the availability of potential energy surfaces of sufficient accuracy to obtain meaningful results. Recently, Tom Miller his group have developed a MOB-ML approach, which provides electronic energies with CCSD(T) accuracy at the expense of a Hartree-Fock calculation. While this represents a significant savings in the computational cost of the calculations, the computational resources required to evaluate the potential surface for, for example, Diffusion Monte Carlo (DMC) calculations remain prohibitive for all but the smallest systems. Likewise, the accurate many-body potentials that have been developed for the evaluation of water clusters and protonated water clusters become computationally prohibitive for DMC studies as the size of the clusters of interest increase. In this talk, we will present a machine-learning-based approach to generate a simple, feed-forward artificial neural network that accurately describes a fitted potential energy surface in the region sampled by the DMC ground state wavefunction. The training and evaluation of these neural network surfaces take advantage of GPU parallelization, which allows for significantly faster parallel potential energy evaluations at each DMC time step when compared to the underlying surface, especially for large ensemble sizes. Applications to water, water dimer and CH5+ will be discussed.

Applying machine learning approaches to studies of ground state properties of fluxional molecules and clusters

ACS Spring Meeting 2021

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