Differentiable modeling and optimization of non-aqueous Li-based battery electrolyte solutions using geometric deep
Shang Zhu1,2, Bharath Ramsundar3, Emil Annevelink1
1Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, USA.
Nature Communications
|October 5, 2024
Summary
We developed DiffMix, a geometric deep learning model, to optimize battery electrolytes. This approach significantly improved ionic conductivity by over 18.8% using robotic experimentation.
Area of Science:
- Materials Science
- Chemical Engineering
- Artificial Intelligence
Background:
- Electrolytes are crucial for next-generation batteries, enabling ion transfer and interface stability.
- Optimizing electrolyte properties for fast charging remains a significant challenge in battery development.
Purpose of the Study:
- To develop a differentiable geometric deep learning (GDL) model, DiffMix, for predicting chemical mixture properties.
- To guide robotic experimentation for optimizing fast-charging battery electrolytes.
Main Methods:
- Extended mixture thermodynamic and transport laws with GDL-learnable physical coefficients.
- Integrated the DiffMix model with a robotic experimentation platform (Clio) for optimization.
- Utilized differentiable optimization based on DiffMix gradients.
Main Results:
- DiffMix demonstrated improved prediction accuracy and robustness compared to purely data-driven models for mixture thermodynamics and ion transport.
- Achieved over 18.8% improvement in electrolyte ionic conductivity within 10 experimental steps.
- Successfully combined GDL, mixture physics, and robotic experimentation for efficient chemical space exploration.
Conclusions:
- DiffMix offers a powerful new predictive modeling method for chemical mixtures.
- The integrated approach enables efficient optimization of battery electrolytes in large chemical spaces.
- This work paves the way for accelerated discovery of advanced battery materials.
