Active machine learning for future battery materials
by Tatu Linnala
The discovery of new materials relies on exploring vast chemical composition spaces. Traditionally, this process requires laborious experimental synthesis or computationally intensive electronic structure simulations, which makes large-scale investigations impractical. In this, artificial intelligence (AI) marks a paradigm shift, fundamentally transforming how materials can be designed and discovered. However, the potential of AI remains underexplored in data-scarce fields, such as the synthesis of new battery materials, where data acquisition is limited by costly experimentation. Developing new battery materials is crucial for advancing sustainable energy technologies, and one promising strategy is to introduce dopants, such as tantalum, aluminum, and calcium, into the cathode material. This has been shown to increase diffusivity, an essential property for battery performance.
Motivated by these challenges and opportunities, my research project employs novel machine learning approaches to accelerate the discovery of new lithium-ion cathode materials. The project will proceed in two phases. In the first phase, suitable Bayesian optimization (BO) methods will be developed and applied to efficiently identify optimal doping levels for high-performance cathodes. To further reduce the need for experimental synthesis during the BO campaign, we plan to utilize existing data from density functional theory simulations to initialize the machine learning model.
In the second phase, the synthesized cathode samples will be characterized using spectroscopic methods to reveal their structural features. Then, machine learning will be used to infer how these structural features influence key material properties, such as diffusivity. These insights on structure-property relations provide further physical understanding and, with it, means of further improving the cathode materials. Ultimately, this research aims to discover novel cathode materials while simultaneously advancing data-driven methodologies for materials discovery.
