The flexible and tunable electronic properties of transition metal (TM) complexes make them promising candidates for functional materials. A near-infinite number of complexes can be defined by varying the ligand field, and virtual high-throughput screening is essential to select leads from this combinatorial design space. This vast chemical space motivates the use of data-driven models to efficiently assess feasible complex properties. We have recently demonstrated the first machine learning (ML) model able to estimate DFT-level electronic and structural properties for TM complexes with good, 3 kcal/mol accuracy, based on simple, metal-centric properties that do not require explicit structural information. By implementing this artificial neural network in our open-source virtual high-throughput screening toolkit, molSimplify, we also enabled the prediction of ground state and geometry to either pre-empt or accelerate computationally-costly first-principles simulations. We also present a new adaptable-basis approach to developing high-dimensional, fully continuous descriptor spaces: these descriptors adapt from fully-nonlocal to highly-local representations of the molecule but still require no structural information. Using feature selection techniques, we extract the most relevant descriptors from this space, identifying the degree of locality or nature (e.g., electronic vs. steric) of atomic molecular properties that most significantly contribute to  properties such as redox potential or spin-state ordering. Using these new continuous descriptors, accuracies in property prediction are also improved, e.g. in kernel ridge regression models: from 3 kcal/mol to sub-kcal/mol in spin-state ordering or from 0.02-0.03 Å to sub-pm in metal-ligand bond length. These down-selected continuous descriptor subsets also provide new definitions of chemical similarity in inorganic space, allowing for interpretation in physical TM complex space of simple arithmetic operations on feature vectors. Finally, we show how these ML models and molecular representations tools are enabling efficient and automated inorganic molecular design.<br/>

Machine learning for inorganic molecular design: Descriptors and similarity in transition metal chemical space

255th National Meeting (2018)

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