Most electronic structure studies in chemistry and biochemistry use methods based on delocalized molecular orbitals because of the computational advantages that the orthogonality of these orbitals provides. The availability of ever increasingly powerful computer systems has led to a renewed interest in the use and new methodological and implementation developments of methods based on non-orthogonal orbitals. Such methods, including valence bond and non-orthogonal configuration interaction, provide more intuitive descriptions of molecular systems for specific classes of applications. An example is charge or energy transfer processes in molecular systems where a description in terms of individually optimized molecules or molecular fragments has significant advantages but results in mutually non-orthogonal fragment wave functions.<br/>One new software development effort for non-orthogonal configuration interaction is GronOR, an international collaboration of the University of Groningen, Oak Ridge National Laboratory, and University Rovira i Virgili. GronOR is designed to construct anti-symmetrized spin-adapted product wave functions for large molecular systems from optimized multi-configuration fragment wave functions and use these as non-orthogonal basis states in subsequent configuration interaction calculations. The fragment wave functions used are complete active space self-consistent field (CASSCF) wave functions, but can, in principle, be generated by any multi-configuration molecular orbital code.<br/>The computationally demanding evaluation of contributions to the Hamiltonian and overlap matrix elements from matrix elements between Slater determinants uses factorization of the second-order co-factor coefficient matrix. Here we present the methodological development and implementation details of GronOR and demonstrate near-linear scalability and high accelerated performance on large, massively parallel, GPU-accelerated computer systems. Example calculations and a comparative benchmark analysis are given for different generations of GPU accelerators as available on Piz Daint at the Swiss Supercomputer Center CSCS, on Summit at the Oak Ridge Leadership Computing Facility OLCF, and on Juwels at the Julich Research Center JFZ.

GronOR: Implementation of non-orthogonal configuration interaction on massively parallel and accelerated supercomputers

ACS Spring Meeting 2021

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