Predicting electronic structure with ANNs

Direct density functional theory (DFT) simulation is the critical bottleneck in high-throughput computational screening for inorganic materials and molecular transition metal complexes. These calculations are computationally costly and the calculated properties are sensitive to the exchange–correlation functional employed. Screening efforts for molecular inorganic systems are complicated by the uncertain spin-state ordering of hypothetical complexes and difficulty in obtaining good initial guess geometries. In this work, we address these challenges by training artificial neural networks (ANNs) to predict quantum-mechanically-derived properties, including spin-state ordering, sensitivity to Hartree–Fock exchange, and spin-state specific bond lengths in transition metal complexes. We use a heuristic set of physical properties that are maximally transferable and do not require precise three-dimensional structural information for prediction. We are able to predict adiabatic spin state splitting energies to within 3 kcal/mol at a range of different Hartree-Fock exchange fractions for a set of different common ligands and first-row transition metals (Cr-Ni). We show superior performance using our ANN compared to kernel ridge regression and support vector regression, and explore the transferability of the trained model to experimentally-characterized transition metal complexes by direct prediction and by extrapolation of GGA results to hybrid functionals. We achieve good performance by interpolation but varied results for direct prediction depending on similarity to the training data and propose a similarity metric to determine if a hypothetical complex will be well-predicted by our model. Our ANNs provide a strategy for screening unexplored metal-ligand chemical space by both property prediction and designing better initial guesses for DFT calculations. 

Check out our recent article in Chemical Science here! Also learn how to directly use our ANN with molSimplify, as introduced in our recent tutorial and also discussed in our recent I&ECR paper.

About Us

The Kulik group focuses on the development and application of new electronic structure methods and atomistic simulations tools in the broad area of catalysis.

Our Interests

We are interested in transition metal chemistry, with applications from biological systems (i.e. enzymes) to nonbiological applications in surface science and molecular catalysis.

Our Focus

A key focus of our group is to understand mechanistic features of complex catalysts and to facilitate and develop tools for computationally driven design.

Contact Us

Questions or comments? Let us know! Contact Dr. Kulik: