Representations and strategies for transferable machine learning improve model performance in chemical discovery

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) for materials discovery is often limited to narrow composition ranges.
• Discovering new materials, especially open-shell transition-metal complexes, is hindered by data scarcity in promising chemical spaces.

Purpose of the Study:

• To develop general and transferable ML models for accelerated discovery across diverse material compositions.
• To address data scarcity challenges in unexplored chemical spaces for transition-metal complexes.

Main Methods:

• Developed an extended graph-based revised autocorrelation (eRAC) representation incorporating group number for isovalent complexes.
• Implemented a transfer learning strategy to seed models trained on abundant data with limited data from new regions.
• Quantified relationships between properties (spin-splitting, ligand dissociation) and periodic trends in isovalent transition-metal complexes.

Main Results:

• The eRAC representation and transfer learning synergistically improved ML model performance.
• Models successfully reordered complex distances to align with periodic table trends, enhancing predictability.
• Demonstrated improved discovery capabilities in regions with limited existing data.

Conclusions:

• The combined eRAC representation and transfer learning strategy effectively accelerates the discovery of transition-metal complexes.
• This approach offers a broadly applicable method for enhancing ML model generalizability across material domains.
• Periodic trends can be leveraged to improve ML model performance in data-scarce chemical spaces.