Benchmarking Neural-Network Variational Monte Carlo (NN-VMC) against the Correlation Consistent Effective Core Potential (ccECP) Dataset

Neural network ansätze for use in quantum Monte Carlo(QMC) represent a recent breakthrough in the application of machine learning approaches to quantum chemistry problems.

Effective core potentials (ECPs) have been developed and employed to reduce the computational cost of calculating target systems of interest, by reducing the number of active electrons in a calculation to only the subset that determine the chemical properties of the system. This work investigates the ability for neural network ansätze to recover the total correlation energy of atomic systems and benchmarks their performance against extrapolated correlated basis set methods such as coupled cluster, as well as fixed-node diffusion Monte Carlo to gauge the relative performance of machine learning driven variational Monte Carlo. In the set of elements from H-Kr, we see closer agreement to the reference energies than previous state of the art quantum Monte Carlo (QMC) methods, with the vast majority of systems falling within chemical accuracy(1kCal/mol) of the predicted total energy. While the computational cost is significant, the accuracy is unmatched for smaller systems unless one does extrapolations to the complete basis set limit. We will also present results on few model diatomic and small metal-clusters, focusing again on the tradeoff between accuracy and cost in obtaining total-energies as well as other one-body observables.