Leveraging Cheminformatics Strategies for Inorganic Discovery: Application to Redox Potential Design

Abstract

Virtual high throughput screening, typically driven by first-principles, density functional theory calculations, has emerged as a powerful tool for the discovery of new materials. Although the computational materials science community has benefited from open source tools for the rapid structure generation, calculation, and analysis of crystalline inorganic materials, software and strategies to address the unique challenges of inorganic complex discovery have not been as widely available. We present a unified view of our recent developments in the open source molSimplify code for inorganic discovery. Building on our previous efforts in the automated generation of highly accurate inorganic molecular structures, first-principles simulation, and property analysis to accelerate high-throughput screening, we have recently incorporated a neural network that both improves structure generation and predicts electronic properties prior to first-principles calculation. We also provide an overview of how multi-million molecule organic libraries can be leveraged for inorganic discovery alongside cheminformatics concepts of molecular diversity in order to efficiently traverse chemical space. We demonstrate all of these tools on the discovery of design rules for octahedral Fe(II/III) redox couples with nitrogen ligands. Over a search of only approximately 40 new molecules, we obtain redox potentials relative to the Fc/Fc + couple ranging from-1 to 4.5 V in aqueous solution. Our new automated correlation analysis reveals heteroatom identity and the degree of structural branching to be key ligand descriptors in determining redox potential. This inorganic discovery toolkit provides a promising approach to advancing transition metal complex design.

Resolving Transition Metal Chemical Space: Feature Selection for Machine Learning and Structure–Property Relationships

“…Cross-validation (CV) scoring, which is unaffected by feature space size changes, will usually produce a minimum for an optimal number of features 80 . We recently used 64 84 , giving the loss function as:…”

“…The redox data set (226 unique structures) is comprised of 41 previously studied 64 Fenitrogen monodentate and bidentate homoleptic complexes and 185 newly generated structures ( Figure 5 and Supporting Information Table S8). The new complexes were obtained by generating combinations of metals (Cr, Mn, Fe, Co) and five small, neutral monodentate ligands (CO, pyridine, water, furan, and methyl isocyanate) with up to two axial ligand types and one equatorial ligand type.…”

“…A key advantage of our geometry-free RACs is that they enable bond length prediction 51 to facilitate accurate structure generation [63][64] . We first evaluate the predictive performance of our full AC set (RAC-155), the proximal subset (PROX-23), and spin-state splitting-selected feature sets (LASSO-28, randF-41, and randF-26) as well as the common subset (RAC-12) for training KRR models on minimum low-spin metal-ligand bond lengths (i.e., min(R L )) in the low-spin, DFT geometry-optimized structures of complexes in the spin-splitting data set.…”

“…All kernel hyperparameter values are provided inSupporting Information Tables S3-S6. Feature selection and model training is carried out on two data sets of single-site octahedral transition metal complexes, which were generated from extension of data collected in previous work51,64 (Figure 5). For both sets, the complexes contain Cr 2+/3+, Mn 2+/3+ , Fe 2+/3+ , Co 2+/3+ , or Ni 2+ first row transition metals.…”

A program for automatically predicting supramolecular aggregates and its application to urea and porphin

“…For complicated aggregate structures, evolutionary algorithms and, possibly, neural networks might be applied for further refining the aggregate candidate set. Periodic boundary conditions might be used at any growth step to check whether experimental properties of extended supramolecular systems are sufficiently well reproduced at that growth step .…”