Large Language Models for High-Throughput Discovery of Novel Metal Oxides for Carbon Capture (Flexible Funding 2024)

Dr Reinaldo Juan Lee Pereira, from Heriot-Watt University, was awarded funding in the UKCCSRC’s Flexible Funding 2024 call to look at “Large Language Models for High-Throughput Discovery of Novel Metal Oxides for Carbon Capture”.

The Challenge of Advanced Materials Discovery

The race for sustainable solutions has scientists worldwide searching for materials that can revolutionize clean energy, carbon capture, and green chemical production. Among promising candidates are perovskite oxides—crystalline materials with the general formula ABO₃ that exhibit remarkable oxygen exchange properties. Their ability to donate and accept oxygen through controlled vacancy formation makes them ideal candidates for chemical looping applications, where the cyclical reduction and oxidation of metal oxides enables efficient chemical conversions without direct contact between reactants.

The Thermodynamics of Chemical Looping

Chemical looping processes rely critically on the oxygen partial pressure (pO₂) exhibited by carrier materials across varying temperatures and redox states. For applications like water-gas shift reactions, materials must exhibit oxygen partial pressures between 10⁻²² and 10⁻¹⁴ bar at operating temperatures around 800°C to achieve conversions exceeding 10%. The challenge is identifying perovskite compositions with the right thermodynamic profile from a vast compositional space—with dozens of potential elements for the A and B sites and variable oxygen stoichiometry, the number of possible candidates reaches into the millions, making conventional screening approaches prohibitively time-consuming.

Leveraging Language Models for Materials Informatics

Our breakthrough applies foundation models from natural language processing to the materials science domain. By fine-tuning pre-trained language models (Phi-4 13B and Llama 3.1 8B) on a curated dataset of DFT-calculated thermodynamic properties, we’ve created a model that can accurately predict formation energies, Gibbs free energies, and oxygen chemical potentials across varying compositions, temperatures, and vacancy concentrations.

Expanding the Frontiers of Computational Materials Science

We’ve procedurally generated approximately 90,000 novel perovskite compositions—exploring combinations far beyond those in existing databases—and rapidly assessed their thermodynamic feasibility for chemical looping applications. While our current focus remains on perovskites, similar approaches could potentially be applied to other functional materials including catalysts, solid-state electrolytes, and high-entropy alloys. This computational screening method demonstrates how AI can significantly accelerate materials discovery, potentially compressing development timelines from years to months for specific applications in energy conversion, carbon capture, and sustainable chemical manufacturing.

Dr. Reinaldo Juan Lee Pereira, Dr. Yongliang Yan, Dr. Ir. Mijndert van der Spek, Dr. Wenting Hu, Dr. Joakim Brorsson, Prof. Tobias Mattisson, Prof. Anders Hellman

Read more on Reinaldo’s Flexible Funding 2024 project page.