Predicting the interface structure

In recent years, artificial intelligence has offered great advances to the field of materials science. With the advent of machine learned potentials, the computationally expensive and chemically accurate methods (such as density functional theory) used to model atomic structures can now be circumvented with these cheaper, yet just as accurate techniques.

This has allowed for hundreds of thousands of bulk materials to be simulated and combed through to identify candidates suitable for next-generation devices with applications in metamaterials, solar cells, batteries and other technologies. The configurational space is massive, with many different potential configurations being potentially real, but the majority are invalid.

This project will involve developing different methods based on a variety of AI techniques to predict the atomic structure of an interface. The work flow will begin from a model with zero data and zero learning. The project will build on existing genetic algorithms, extending to utilising advances, such as message passing neural networks, to more efficiently predict atomic configurations and their favourabilities.

The approach will involve a recursive approach of generating data and feeding it back into the method to update itself, thus learning on-the-fly. By cross pollinating between these two methods, we intend to develop a method that can predict these structures in sensible timescales for minimal cost. This will involve developing custom descriptor approaches which adapt to each individual "system." In the latter part of the project, dependent on the success of the first part, we will develop a second more holistic approach which uses the data of all the different systems to explore trends.

New methodologies will need to be developed to overcome challenges, such as small dataset sizes, fine-tuning far from training data, and need to adapt to individual environments. Developing such a capability would allow the exploration of new metamaterials, and allow the creation of new technologies.