Materials design using machine learning
The complex task of developing future materials, built on the stable crystal forms of new molecules with outstanding properties, relies on the speed and power only available through supercomputing. By using highly specialised codes, researchers can get supercomputers to tackle this challenge pragmatically rather than by making incomplete ‘educated guesses’. Dr Jack Yang from the School of Material Science and Engineering at the University of New South Wales is using NCI’s supercomputer to incorporate machine learning into his own materials predictions, speeding up and simplifying the process even further.
In a paper published in the journal Chemistry of Materials, Dr Yang and his co-authors showed one of the largest crystal structure predictions for molecular organic semiconductors in a single paper. They used structure predictions and machine learning techniques that ran on NCI’s supercomputer to identify and then classify the crystalline arrangements of 28 different precursors for molecular semiconductors. From here, they will be able to further generalise about how the materials’ properties correlate with their complex three-dimensional crystal structures.
“At UNSW we are trying to develop a centre of expertise in nanomaterials and molecular crystals. We are bringing chemistry, physics and materials science together in order to gain deeper insights from high-throughput atomistic simulations. We will use these to guide the experimental work that ultimately bring these new materials into reality.”
“Apart from doing the fundamental science, it takes a lot of expertise and time to develop the research software and maintain the data that we need, an aspect that has been largely overlooked in the community,” Dr Yang says.
“Developing research software in the age of big-data really needs a different mindset for those of us who are not from a software engineering background.”
“Using a supercomputer to generate crystal structures is easy, but understanding all those structures is hard, and particularly why some structures are more stable than others. We are developing better ways of using supercomputing, and machine learning tools, to analyse which of the theoretical structures to investigate further in the lab.”
Understanding how the properties of materials come about, and through which molecular arrangements, will help future researchers more quickly analyse their own experimental structures. Researchers will be able to rank promising molecules as synthetic targets based on their potential as useful semiconductor materials, pharmaceutical drugs or photocatalysts. The prediction of new structures, as well as their analysis, continues to rely on the power of advancing supercomputer and big data technology, which NCI is providing to the nation’s researchers.