The tunable electronic structure properties of transition metal complexes make them attractive for application in catalysis and energy conversion and storage but also make the exploration of their “neighborhood” in vast chemical space daunting. The high computational cost of obtaining first-principles results and the nearly infinite design space spanned by ligand and metal chemistries mandates cheap techniques to obtain property estimates for screening and rational design. However, many well-established concepts of molecular similarity for organics (e.g. fingerprint methods) correlate poorly with transition metal properties or depend on expensive-to-compute (i.e., with first-principles density functional theory, DFT) optimal geometries, making it challenging to develop quantitative structure-property relationships (QSPRs), e.g. predictive machine learning models. We have recently used descriptors based on heuristic physical properties, such as properties of the metal and adjacent atoms, and connectivity in the molecular graph, to train an artificial neural network (ANN) to predict DFT properties of transition metal complexes to 3 kcal/mol accuracy, outperforming more commonly employed descriptors suitable for organic chemistry by an order of magnitude on our test set. We now introduce new, flexible, continuous descriptors of transition metal complexes based on modified autocorrelation functions which separately encode information from the first and subsequent coordination shells, along with overall ligand and complex information, providing customized descriptors suitable across inorganic chemistry applications. The result is an easily-computed embedding from chemical design space into a high-dimensional vector space, which provides a suitable setting for defining similarity across varying ligands environments and metal centers. Using numerous established techniques in feature selection, we identify strategies to extract the most important variables from this high-dimensional space to yield optimal subsets for machine learning QSPRs on a few hundred unique complexes. These new models match or exceed performance from our previous descriptor set. We extract chemical insight from our feature space selection results using principal component analysis and self-organizing maps, identifying how different choices of descriptor spaces alter the ‘chemical resolution’ of the embedding.

Mapping transition metal chemical space with continuous descriptors – feature selection and implications for machine learning models

254th National Meeting (2017)

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