Accelerating discovery of novel CO2 capture materials using AI and machine learning
Project duration: July 2023–March 2025
Dr Aaron Thornton
Ravichandar Babarao, Tu Le and Ibrahim Barry Orhan.
Direct air capture (DAC) requires a liquid or solid material that selectively adsorbs carbon dioxide (CO2) from air. Discovering next-generation DAC materials that capture more CO2, more efficiently, will help reduce the size and energy requirements of what is a mega-scale operation. Unfortunately, discovering new materials is a slow process. Given the urgency and the need for scaled negative emission technologies, we need to accelerate progress.
Discovering new materials is seen as a current hurdle in optimising DAC, and this knowledge gap needs to be addressed. If human input can be minimised, then optimised research outcomes can be reached more rapidly and efficiently. As recent research shows, AI and machine learning can offer a faster approach than human endeavour, and it’s already been used for single compound discoveries. But can we teach chemical intuition to AI to enable it to identify next-generation biological and composite porous materials for DAC? This would mean a fundamental shift in materials discovery with potential research impact across many areas of science including drug discovery, solar cells and batteries, and biomedical materials.
There is currently a large gap in chemical information for AI. For example, in order to identify novel materials AI first needs to learn what makes a good DAC material, and the information needs to be in language that AI can understand and utilise. By combining machine learning and molecular simulation we aim to fill this knowledge gap and use AI to predict, model and discover new composite materials for CO2 capture from air.
We’ll do this by expanding the materials datasets and developing a new set of features that AI can understand. Then we’ll adapt the algorithms to biological molecules and other porous materials and apply that capability. Results will be used to develop a work plan for production of new materials. The prime value of the research lies in significantly reducing the time that’s required to discover new materials. Given the infinite combinations of chemical building blocks, and the time and cost of lab-based experiments, adding an advanced virtual screening layer prior to experimentation will accelerate the discovery of new materials for DAC and, potentially, many other applications that involve the capture of small molecules such as ammonia recovery from hydrogen production and methane from mines.
The greatest challenge is developing the chemical language that AI can understand and utilise, and translating this into machine learning features. Other technical risks of this approach relate to the availability of material datasets, predictivesimulation tools, machine learning models and molecular simulation workflows, as well as computational resources. To overcome these challenges we’ll use the knowledge and insight gained in a previous study, which identified several new candidate materials from an existing database.
Success largely depends on the quality of the input, and we bring a deep understanding of the underlying CO2 capture mechanisms, the adsorption process, chemistry of new materials, activity of proteins, advanced knowledge of machine learning, and high-performance computing resources to meet the challenge.
Young, J., Mcilwaine, F., Smit, B., Garcia, S., & Van der Spek, M. (2023). Process-informed adsorbent design guidelines for direct air capture. Chemical Engineering Journal, 456, 141035.
Findley, J. M., & Sholl, D. S. (2021). Computational screening of MOFs and zeolites for direct air capture of carbon dioxide under humid conditions. The Journal of Physical Chemistry C, 125(44), 24630-24639.