Resolving Transition Metal Chemical Space: Feature Selection for Machine Learning and Structure-Property...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Informatics

Background:

• Machine learning (ML) models are crucial for accelerating chemical discovery by predicting quantum mechanical properties.
• Accurate ML predictions for transition metal chemistry are challenging due to high computational costs and limited training data.
• Molecular representation is a critical factor influencing ML model accuracy in these scenarios.

Purpose of the Study:

• To introduce novel molecular descriptors, revised autocorrelation functions (RACs), designed for improved ML accuracy in inorganic chemistry.
• To adapt standard autocorrelation functions (ACs) for better applicability to the complexities of inorganic molecules.
• To demonstrate the effectiveness of RACs in predicting key chemical properties for both organic and inorganic systems.

Main Methods:

• Developed a series of revised autocorrelation functions (RACs) by modifying standard ACs to encode atomic property relationships on molecular graphs.
• Applied RACs to ML models for predicting atomization energies in organic molecules and spin-state splitting in inorganic systems.
• Compared RAC performance against existing topological descriptors and whole-molecule structural features.
• Investigated systematic feature selection methods (univariate filtering, recursive feature elimination, random forest, LASSO) to optimize descriptor subsets.

Main Results:

• RACs achieved significantly lower mean unsigned errors (MUEs) for atomization energies (as low as 6 kcal/mol) on organic molecules compared to other descriptors.
• For inorganic chemistry, RACs yielded exceptionally low MUEs (1 kcal/mol) for spin-state splitting, outperforming traditional features by 15-20 times.
• Optimized subsets of RACs, selected via random forest or LASSO, achieved sub- to 1 kcal/mol MUEs for spin-splitting.
• Demonstrated good transferability of RACs to predict metal-ligand bond lengths (0.004-5 Å MUE) and redox potentials (0.2-0.3 eV MUE).

Conclusions:

• Revised autocorrelation functions (RACs) represent a significant advancement in molecular representation for ML in chemistry.
• RACs provide superior predictive accuracy for critical inorganic properties like spin-state splitting, even with reduced feature sets.
• The choice of descriptors (local electronic vs. distal steric) is crucial for accurately predicting different chemical properties, highlighting the versatility of RACs.