Investigations into the physical nature of complex phenomena increasingly depend upon the unified investigation of theoretical and experimental methods.  However, as the size and complexity of these systems increases, quantum chemical simulation methods become increasingly costly, particularly in studies pertaining to extended solid state systems. Consequently, methods which enhance numerical accuracy while minimizing the corresponding computational scaling with respect to system size are highly desirable.  Multi-agent consensus equilibrium is a criterion which may be used with convex sets of functions to fuse multiple sets of data, under the constraint that each agent describes the system in the same dimensionality. <br/> <br/>In the presented work, MACE is utilized to fuse the output of multiple methods in computational chemistry to a single combined result. Preliminary studies integrate results computed from PM3, Hartree-Fock/6-31+G**, MP2/6-31+G**, B3LYP/6-31+G**, and Hartree-Fock/cc-pVDZ methods, each computed in parallel. The MACE result represents the intersection of manifolds corresponding to each agent and consequently the recovered result is not equivalent to the arithmetic average of each individual model.  Furthermore, the intrinsic uncertainty within each agent (e.g., the slope of the potential energy surface in a geometry optimization) provides unsupervised feedback into the MACE criterion, dictating the equilibrium state. Preliminary MACE results optimizing the geometry of water and other small molecules demonstrating greater accuracy than gradient descent upon the individual agents will be presented.  The capacity for integrating experimental measurements as agents and the greater applicability of the method discussed. Application of this data-science based approach to <i>ab initio</i> chemistry aims to improve the accuracy of quantum chemistry without increasing computation time beyond the agent with the largest scaling.

Multiagent consensus equilibrium (MACE) for addressing the scaling challenges of computational chemistry

2019 Fall National Meeting

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