Abstract

Obtaining quantitatively accurate potential energy surfaces (PESs) for molecular systems (with many dimensions and complex electronic structure) is challenging due to the increased computational cost related to simply getting the training data. It then becomes important to efficiently select, a priori, the molecular geometries at which one wants to calculate this expensive data, and to use it efficiently.

We apply Gaussian processes to learn the energy surfaces as well as the corrections, using delta-learning. We present various ways of using test sets from electronic structure calculations, as well as the intrinsic Gaussian processes covariance functions, to generate optimal training sets for further PES generation. We present results for a water dimer proton exchange PESs from DFT data (PBE//aug-cc-pVDZ) to coupled cluster ( CCSD -F1,2/aug-cc-pVTZ) accuracy along all 12 dimensions. The same principles are also used on small molecules inside fullerenes at RIMP2 //cc-pVDZ on the cage and RIMP2 //cc-pVQZ on the internal molecule accuracies for the internal degrees of freedom.