Why is this research needed?

This research will leverage large language models (LLMs), a class of machine learning models, to predict and generate novel perovskite oxides for carbon capture processes using chemical looping.

There is significant interest in applying metal oxides for power, hydrogen, syngas, alkene, and epoxy production since pure and capture-ready CO2 streams can be produced via chemical looping process (i.e., a type of process intensification), thus circumventing the need for costly and energy intensive CO2 separation processes (e.g. amine sorbents). Though for each of these applications, a different metal oxide with suitable thermodynamic and kinetic properties must be selected. Perovskite oxides are a class of materials which are tuneable, composable, thermally and mechanically stable, possess fast oxygen mobility, and have high selectivity’s to gases, making them ideal candidates for this purpose. Experimental and theoretical methods for testing and discovering perovskite oxides, though necessary, are resource intensive methods. The advent of artificial intelligence and machine-learning (ML) methods has reduced this effort by utilizing a small dataset (from experimental or theoretical work) to predict a much wider number of hypothetical perovskites in other applications (e.g. Solar cells, high-entropy alloys). Current materials databases do not possess a sufficiently large number of perovskite oxides to train a conventional ML model for chemical looping applications, furthermore, the accuracy of properties provided by databases such as the materials project or OQMD are not sufficient to appropriately predict the desired material properties. Of note is that machine learning models based on LLMs have shown accurate predictions for material properties using smaller datasets (e.g., DFT computed data or experimental work) due their pre-embedded understanding of chemistry and material properties. By allowing high-quality yet small datasets to be used, discovery of novel perovskite oxides with accurate properties can be achieved.

What is this research investigating?

1. Construct a large language machine learning model based on DFT computed datasets containing properties of perovskite oxides with pre-defined metal compositions and oxygen non-stoichiometry.
2. Predict properties of existing crystalline perovskite materials with similar or better accuracies compared to materials databases (<0.1 eV/atom).
3. Generate new perovskite compositions with the desired properties.

What does the research hope to achieve?

The proposed work will advance applications of LLMs for materials screening and is not limited to screening perovskite oxides for chemical looping but also for solid oxide fuel cells, catalysis, energy storage, and sensors. In future work, perovskites screened from this work can be synthesized and tested using equipment available at Newcastle University or Chalmers University. Datasets generated from this work will be published in an open-access repository such as GitHub, and the resulting model is expected to be integrated into a pre-existing workflow for screening metal oxides (published on GitHub). This work will lead to at least one publication in an open-access and peer-reviewed journal. Results from this work will be disseminated at UKCCSRC events (Spring/Autumn conference), international conferences (CHEMUK2025, STEM for Britain), and on social media accounts (RCCS group website, MatCoRe group website, LinkedIn). Participation in AI Hackathons (NFDI4) will be used to share experiences from the project (with potential for a conference proceeding or journal article).