Chemistry-Informed Machine Learning for Polymer Electrolyte Discovery

Area of Science:

• Materials Science
• Electrochemistry
• Computational Chemistry

Background:

• Solid polymer electrolytes (SPEs) offer enhanced safety and higher energy density for lithium-ion batteries.
• SPEs currently exhibit lower ionic conductivity compared to liquid and solid ceramic electrolytes, hindering their practical application.
• Accelerating the discovery of SPEs with high ionic conductivity is crucial for advancing battery technology.

Purpose of the Study:

• To develop a chemistry-informed machine learning model for accurate prediction of SPE ionic conductivity.
• To overcome the limitations of existing models by incorporating temperature dependence.
• To identify promising SPE formulations for next-generation batteries.

Main Methods:

• Developed a machine learning model integrating the Arrhenius equation into a message passing neural network.
• Trained the model on ionic conductivity data from hundreds of experimental publications.
• Predicted ionic conductivity for thousands of candidate SPE formulations.

Main Results:

• The chemistry-informed model significantly improved prediction accuracy by encoding temperature dependence.
• Identified promising candidate SPE formulations with high ionic conductivity.
• Demonstrated the model's utility in predicting conductivity for various anions and polymer hosts.

Conclusions:

• Chemistry-informed machine learning models can effectively accelerate the discovery of advanced materials like SPEs.
• Incorporating physical principles, such as temperature dependence, enhances model accuracy, especially with limited data.
• This approach facilitates the identification of novel SPEs for safer and more powerful lithium-ion batteries.