Machine learning for potential energy surfaces: An extensive database and assessment of methods

Abstract

On the basis of a new extensive database constructed for the purpose, we assess various Machine Learning (ML) algorithms to predict energies in the framework of potential energy surface (PES) construction and discuss black box character, robustness, and efficiency. The database for training ML algorithms in energy predictions based on the molecular structure contains SCF, RI-MP2, RI-MP2-F12, and CCSD(F12*)(T) data for around 10.5 × 106 configurations of 15 small molecules. The electronic energies as function of molecular structure are computed from both static and iteratively refined grids in the context of automized PES construction for anharmonic vibrational computations within the n-mode expansion. We explore the performance of a range of algorithms including Gaussian Process Regression (GPR), Kernel Ridge Regression, Support Vector Regression, and Neural Networks (NNs). We also explore methods related to GPR such as sparse Gaussian Process Regression, Gaussian process Markov Chains, and Sparse Gaussian Process Markov Chains. For NNs, we report some explorations of architecture, activation functions, and numerical settings. Different delta-learning strategies are considered, and the use of delta learning targeting CCSD(F12*)(T) predictions using, for example, RI-MP2 combined with machine learned CCSD(F12*)(T)-RI-MP2 differences is found to be an attractive option.