Navigating Transition-Metal Chemical Space: Artificial Intelligence for First-Principles Design.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning Applications

Background:

• Open-shell transition-metal complexes offer diverse functionalities but pose challenges for conventional computational modeling due to complex electronic structures.
• Density functional theory (DFT), while a workhorse, is computationally expensive and prone to inaccuracies for localized d-electrons, limiting its use in exploring vast transition-metal chemical space.
• Existing methods for small organic molecules are not directly applicable to the unique challenges presented by transition-metal complexes.

Purpose of the Study:

• To develop and apply machine learning (ML) models for efficient prediction of properties in open-shell transition-metal complexes.
• To explore the vast chemical space of transition-metal complexes for the design of novel functional materials and catalysts.
• To establish robust methods for property prediction, design rule encapsulation, and uncertainty quantification in ML models for this domain.

Main Methods:

• Development of novel machine learning models, including artificial neural networks and kernel ridge regression, coupled with tailored representations for transition-metal complexes.
• Application of ML models to predict geometric, spin state, and redox potential properties, subsequently extended to catalysis and gas separation properties.
• Implementation of distance-based approaches for model uncertainty quantification and domain applicability assessment.

Main Results:

• Machine learning models accurately predict key properties of transition-metal complexes, overcoming limitations of traditional DFT.
• Interpretation of ML models reveals essential structure-property relationships and design rules.
• ML enables rapid assessment of millions of compounds, identifying promising new material candidates in weeks.

Conclusions:

• Machine learning provides a powerful and efficient engine for exploring the complex chemical space of transition-metal complexes.
• This approach significantly accelerates the discovery and design of new functional materials and catalysts.
• ML-driven exploration, combined with uncertainty quantification, offers a robust strategy for tackling multi-objective design problems.