Improvements in computing power and algorithmic advances have made atomistic simulation, particularly with density functional theory (DFT), an essential tool in designing materials to meet many pressing technological challenges. However, first-principles simulation remains too costly to address truly high-dimensional design challenges and this problem is especially acute for open shell transition metal (TM) complexes, a class of materials with applications in homogeneous catalysis and molecular electronics. We address this challenge by supplementing DFT calculations with machine-learned ‘surrogate models’, artificial neural networks (ANNs) which can provide no-cost estimations of TM complex properties at accuracies that are comparable with baseline DFT uncertainty for these systems. By careful consideration of model uncertainty, we developed an evolutionary design strategy that is able to search large candidate spaces for TM complexes with targeted spin state ordering and frontier orbital properties. The majority of our leads are validated at the DFT level, representing at least an order-of-magnitude acceleration over DFT alone. We also train on calculations using different fractions of exact exchange and hence are able to provide estimates of properties in a DFT functional-dependent way, addressing uncertainties in the application of DFT to these systems. Finally, we are able to improve DFT screening directly by predicting DFT optimized geometries and using them initialize calculations. Taken together, this work represents substantial progress towards the goal of accelerating the <i>in silico</i> design of novel inorganic complexes with targeted properties, combining expensive quantum simulations and cheap machine-learned surrogate models. All models and code developed are open-source and distributed through GitHub via https://github.com/hjkgrp.<br/>

Multifidelity methods for the design of transition metal complexes

2019 Fall National Meeting

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