GPU-Accelerated Multireference Methods for Polynuclear Transition Metal Complexes

Computational modelling of molecules with multiple transition metals having strongly correlated electrons is an ever-present and a diffcult challenge in the field of theoretical chemistry. Conventional multireference wave function methods have been used successfully for strongly correlated systems but the high cost associated with them limits their usage for polynuclear transition metals. Practically, a numerical efficient method using modern computing architecture is needed to reduce the time-to-solution.

This dissertation focuses on development and application of GPU-accelerated chemically-guided fragmentation based multireference methods. The central exploration is with the localized active space self-consistent field (LASSCF) method. The method models locally correlated centers in spatially separated regions of a molecule. While the method scales better than the conventional complete active space (CAS), it can give qualitatively incorrect wave function in strongly correlated fragments. The LAS-state interaction method adds back correlation explicitly. This work scales the method to highly challenging systems.

To make the method numerically efficient, we have leveraged the modern GPU-accelerated computing architecture. We have ensured the usage of portable programming practices to retain performance across various hardware and software ecosystems. With this, we have studied the [Fe4S4(SMe)4]2− cluster and tried to construct simple models for the low-lying electronic states.

This work also expands our abilities to model challenging polynuclear transition metal complexes and obtain qualitatively correct wave functions with high numerical speed. As a byproduct of this work, we also have access to order of magnitude faster conventional methods like HF, DFT and CASSCF