Resolving transition metal chemical space: feature selection for machine learning and structure-property relationships

TL;DR

This paper introduces revised autocorrelation functions (RACs) for machine learning in transition metal chemistry, demonstrating improved accuracy in predicting properties like spin-state splitting, bond lengths, and redox potentials with smaller feature sets.

Contribution

The study develops RACs tailored for inorganic chemistry and systematically compares feature selection methods, achieving high accuracy with reduced feature sets for various properties.

Findings

RACs outperform standard topological descriptors in organic molecules.

ML models with selected features achieve 1 kcal/mol accuracy in spin-state splitting.

Feature importance varies with property, highlighting local electronic and steric effects.

Abstract

Machine learning (ML) of quantum mechanical properties shows promise for accelerating chemical discovery. For transition metal chemistry 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. We introduce a series of revised autocorrelation functions (RACs) that encode relationships between the heuristic atomic properties (e.g., size, connectivity, and electronegativity) on a molecular graph. We alter the starting point, scope, and nature of the quantities evaluated in standard ACs to make these RACs amenable to inorganic chemistry. On an organic molecule set, we first demonstrate superior standard AC performance to other presently-available topological descriptors for ML model training, with mean unsigned errors (MUEs) for atomization energies on set-aside test molecules as low as 6 kcal/mol. For inorganic chemistry, our RACs yield 1 kcal/mol ML MUEs on set-aside test molecules in spin-state splitting in comparison to 15-20x higher errors from feature sets that encode whole-molecule structural information. Systematic feature selection methods including univariate filtering, recursive feature elimination, and direct optimization (e.g., random forest and LASSO) are compared. Random-forest- or LASSO-selected subsets 4-5x smaller than RAC-155 produce sub- to 1-kcal/mol spin-splitting MUEs, with good transferability to metal-ligand bond length prediction (0.004-5 {\AA} MUE) and redox potential on a smaller data set (0.2-0.3 eV MUE). Evaluation of feature selection results across property sets reveals the relative importance of local, electronic descriptors (e.g., electronegativity, atomic number) in spin-splitting and distal, steric effects in redox potential and bond lengths.