Navigating Ternary Doping in Li-ion Cathodes With Closed-Loop Multi-Objective Bayesian Optimization

Area of Science:

• Materials Science
• Electrochemistry
• Machine Learning

Background:

• Optimizing secondary battery materials requires exploring complex composition spaces.
• Previous methods like grid search are inefficient for multi-component systems.
• High-throughput experimentation is valuable but faces scalability challenges with increasing complexity.

Purpose of the Study:

• To develop a closed-loop, multi-objective machine learning approach for efficient battery material discovery.
• To navigate a vast compositional space of approximately 14 million unique combinations.
• To simultaneously optimize multiple electrochemical properties beyond just energy density.

Main Methods:

• Utilized a set transformer pretrained on the Materials Project database for feature extraction.
• Employed a multi-task Gaussian process model for predicting electrochemical properties.
• Integrated machine learning with a high-throughput workflow using active learning over 3 rounds.

Main Results:

• Successfully optimized four key electrochemical properties simultaneously using a small number of samples (125 random, 63 predicted).
• Identified a LiCoPO4 composition that increased the composite figure of merit up to five times compared to the undoped system.
• Demonstrated an end-to-end workflow for accelerated battery materials design.

Conclusions:

• The developed machine learning approach significantly enhances the efficiency of secondary battery material discovery.
• This methodology enables simultaneous optimization of multiple critical electrochemical properties.
• The workflow is poised to accelerate the field of autonomous materials discovery for advanced batteries.