Using Machine-Learning to Create Predictive Material Property Models

Chris Wolverton
Northwestern University

Rational, data-driven materials discovery would be an immense boon for research and development, making these efforts far faster and cheaper. In such a paradigm, computer models trained to find patterns in massive chemical datasets would rapidly scan compositions and systematically identify attractive candidates for technological applications, such as new batteries, lightweight alloys, solar cells, and so on. Here, we present several examples of our work on developing machine learning (ML) methods capable of creating predictive models using a diverse range of materials data. As input training data, we demonstrate ML on both large computational datasets of density functional theory calculations, as implemented in the Open Quantum Materials Database (oqmd.org), and also experimental databases of materials properties. We construct ML models using a large and chemically diverse list of attributes, which we demonstrate are suitable for describing a range of properties, and a novel method for partitioning the data set into groups of similar materials in order to boost the predictive accuracy of the ML models. Our results demonstrate how ML can be used as an effective tool to automatically learn intuitive design rules, predict diverse properties of crystalline and amorphous materials, such as formation energy, specific volume, band gap energy, and glass-forming ability, and accelerate combinatorial searches.

Back to Workshop I: Machine Learning Meets Many-Particle Problems