High-Conductivity Electrolytes Screened Using Fragment- and Composition-Aware Deep Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Electrochemistry

Background:

• Global demand for long-duration energy storage is increasing due to renewable energy growth.
• Rechargeable batteries are crucial for grid stability, with electrolyte performance being key.
• Current electrolyte design relies on slow, resource-intensive trial-and-error methods.

Purpose of the Study:

• To develop a novel deep learning framework for accurate and interpretable prediction of ionic conductivity in battery electrolytes.
• To overcome limitations of existing machine learning models that use coarse representations and neglect chemical details.

Main Methods:

• A hierarchical deep learning framework integrating intermolecular and intramolecular attributions.
• Decomposition of formulations into molecules, functional units, and ratios.
• Generation of mixture-invariant embeddings using physicochemical descriptors and salt identity.

Main Results:

• High accuracy in predicting ionic conductivity on benchmark datasets.
• Enabled large-scale virtual screening of electrolyte formulations.
• Provided chemically interpretable insights into fragment-level contributions and ratio effects.

Conclusions:

• The framework facilitates data-driven and interpretable electrolyte design for lithium batteries.
• It offers a more efficient alternative to traditional design methods.
• The approach is generalizable to other materials formulation challenges.