Alumnus Taps Machine Learning to Find New Materials
The search for effective materials to store renewable energy could accelerate with a technique a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) alumnus has developed to apply machine learning to the task.
Zachary Ulissi, an assistant professor at Carnegie Mellon University (CMU), and doctoral candidate Kevin Tran describe the technique in a recent paper published in Nature Catalysis. They say their automated screening method uses machine learning and optimization to guide molecular models that predict the performance of new catalysts that are useful for reducing carbon dioxide or splitting water into hydrogen and oxygen.
As Ulissi, a fellow from 2010 to 2014, explains in this story from the CMU College of Engineering, finding these useful materials usually takes a combination of intuition and hours of routine calculations that may yield little if anything. The method he devised with Tran searches a database of millions of adsorption sites – places another element could adhere – on thousands of intermetallics, a particular metal class. With that search, it builds a machine-learning model to predict what site it should run calculations on. Those computations reveal more about the intermetallic site and are used to retrain the model. Each time the loop repeats, it finds better and more interesting materials, discounting those that wouldn’t make good catalysts while directing researchers to more promising options.
In the Nature Catalysis paper, Tran and Ulissi list dozens of materials the technique suggests would work well for carbon dioxide reduction and water splitting. The method can study hundreds of proposed materials per day; a human could only analyze 10 to 20 per week.
Ulissi and DOE CSGF alumna Brenda Rubenstein, a Brown University chemistry professor, also are participants in a recently announced project to develop advanced software capable of designing chemicals and processes for energy production and other potential applications. The project, led by Brown’s Andrew Peterson, will work on approaches to bridge time scale differences when computing chemical systems on future computing systems capable of a quintillion (1018) scientific calculations per second.
Learn more about Ulissi’s CMU work in this video.