High-throughput experimentation and machine learning for materials discovery

Ichiro Takeuchi
University of Maryland

We have developed a comprehensive high-throughput methodology for rapid screening of large compositional phase space in search of new materials with enhanced physical properties. Thin-film combinatorial libraries are characterized quantitatively using a variety of rapid measurement tools including synchrotron x-ray diffraction. In order to streamline the analysis of large volumes of data from the libraries, machine learning is used to quickly perform clustering of structural phases and composition regions with similar properties across phase diagrams. In a mode called Integrated Materials Discovery Engine, we combine high-throughput experimentation with theoretical approaches with multiple feedback between the two in order to rapidly converge on optimized materials. Incorporation of databases such as ICSD and AFLOW for rapid cross-referencing and validation is a crucial part of our analysis strategy. To date, we have successfully used the approach to discover new magnetic materials, shape memory alloys, piezoelectric materials, and superconductors. I will also discuss our latest effort where we are employing active learning, a form of machine learning, to determine the sequence of individual experiments in order to maximize attainable knowledge from each point, further speed up the overall process, and minimize the amount of experimental resources. This work is carried out in collaboration with Gilad Kusne, Stefano Curtarolo, Valentin Stanev, Apurva Mehta, and Tieren Gao and is funded by ONR, AFOSR, and NIST.

Presentation (PDF File)

Back to Workshop I: Optimization and Optimal Control for Complex Energy and Property Landscapes