Machine learning (ML) of quantum mechanical properties shows promise for accelerating chemical discovery. For transition metal catalysis where accurate calculations are computationally costly and available training data sets are small, the molecular representation becomes a critical ingredient in ML model predictive accuracy. By exploiting locality in transition metal chemistry, we recently trained the first artificial neural network (ANN) for metal-organic bonding in transition metal complexes on a series of metal-proximal descriptors that required no explicit structural information. This ANN demonstrated good performance on set-aside test molecules (e.g., 3 kcal/mol mean unsigned error on spin-state splitting and 0.02 Å on bond lengths) and outperformed conventional whole-molecule descriptors (e.g., the Coulomb matrix) widely used in organic machine learning and demonstrated enhanced transferability for discovery. I will discuss our strategies for ANN-driven chemical discovery at a fraction of the computational cost of first principles simulation with evolutionary algorithms that balance expected compound fitness with expected ANN prediction accuracy. I will describe our new “complete basis” approach to descriptor development, in which heuristic properties of individual atoms such as electronegativity, size, and connectivity are correlated (i.e., in a sum of products of atomic quantities) at adaptable length scales across the molecular graph. This approach still does not require structural information prior to ML prediction. Importantly, when paired with rigorous feature selection methods across property sets, selected descriptor subsets from this complete basis reveal the length-scale and nature of the critical atomic properties such as the importance of local, electronic descriptors (e.g., electronegativity, atomic number) in spin-splitting and distal, steric effects in redox potential and bond lengths. These new descriptors also reduce prediction energetic and geometric errors to sub-kcal/mol and sub-pm, respectively, in trained ML models. We have implemented descriptor evaluation and ML models in our open-source inorganic molecular discovery toolkit, molSimplify. I will describe some of our recent work in automating molecular design workflows in molSimplify with both machine-learning- and first-principles-driven design of C-H activation catalysts and the optimization of redox potential and spin-state ordering in single-site catalysts.

Choosing the right chemical representation for machine-learning-accelerated discovery and design in transition metal catalysis

255th National Meeting (2018)

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