Formulation Graphs for Mapping Structure-Composition of Battery Electrolytes to Device Performance.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Discovering new formulations, like advanced materials, is challenging.
• Current methods rely heavily on laboratory experiments for formulation optimization.
• High-throughput screening of components accelerates compound discovery but not formulation selection.

Purpose of the Study:

• To develop a deep learning model for predicting liquid formulation properties.
• To establish a computational method for mapping structure-composition relationships to formulation performance.
• To accelerate the discovery and development of novel formulations.

Main Methods:

• Developed the Formulation Graph Convolution Network (F-GCN) model.
• Utilized parallel Graph Convolutional Networks (GCNs) to featurize formulation constituents.
• Integrated constituent descriptors, scaled by molar percentage, into a combined formulation descriptor.
• Employed knowledge transfer with HOMO-LUMO and electric moment properties for enhanced molecular descriptors.

Main Results:

• The F-GCN model accurately predicts performance metrics for battery electrolytes.
• Achieved the lowest reported errors in predicting Coulombic efficiency (CE) and specific capacity.
• Demonstrated efficacy on two distinct datasets for battery electrolyte formulations.
• The best-performing model utilized molecular graph descriptors informed by electronic properties.

Conclusions:

• The F-GCN model offers a powerful computational tool for formulation discovery.
• This deep learning approach significantly reduces reliance on experimental screening for formulations.
• Enables faster development of high-performance materials, particularly in energy storage applications.