Autonomous Combinatorial Experimentation for Streamlined Materials Discovery 

Ichiro Takechi¹

    ¹University of Maryland

Abstract: We have established the thin film based combinatorial high-throughput approach for rapid investigation of novel functional materials [1]. Over the years, we have used it to make a number of interesting and important materials discoveries. They include new lead-free morphotropic phase boundary piezoelectric materials, shape memory alloys with long fatigue life, heterogeneous magnetostrictive materials, and high-performance phase change memory nanocomposite materials [2]. 

     The initial challenges associated with combinatorial experimentation were 1) how to reproducibly make combinatorial libraries which encompass large compositional landscapes and 2) how to accurately and quickly measure various physical properties across libraries. By mid 2000s, the main challenge had shifted to how to handle the large amount of data which are churned out of combinatorial experimentation. We have then started incorporating various machine learning approaches in order to quickly analyze and interpret the multidimensional materials data, not just for identifying new compositions of materials with enhanced physical properties, but also for delineating the composition-structure-property relationships which underpin the physical mechanisms giving rise to enhanced physical properties. This was one way the community dedicated to combinatorial and high-throughput experimentation had learned to embrace materials informatics. 

     With the advent of the Materials Genome Initiative in 2010s, there was a major shift in the materials exploration trends in the materials science community at large. The age of high-throughput computational materials science had arrived, and large databases of materials properties based on computational predictions began to emerge. While initially the main focus of the Materials Genome Initiative was squarely on computational work, the community soon realized that the experimental high-throughput strategies are necessarily an integral part of the Initiative. Materials informatics has now become a household name helped in part by the high-throughput materials science community. 

     As a branch of machine learning, active learning has attracted much attention recently. It can effectively help navigate experimental sequences in materials research. We are actively incorporating active learning in screening of combinatorial libraries of functional materials. The array format with which samples with different compositions are laid out on combinatorial libraries is particularly conducive to active learning driven autonomous experimentation. For some physical properties, each characterization/measurement requires time/resources long/large enough that "high"-throughput measurement is not possible. Examples include detection of martensitic transformation and superconducting transitions in thin film libraries. By incorporating active learning into the protocol of combinatorial characterization, we can streamline the measurement and the analysis process substantially. We have developed several schemes to carry out autonomous combinatorial experimentation, where Bayesian active learning algorithms dictate the sequence of experiments [2]. I will illustrate how implementation of these schemes allows us to substantially reduce the overall time and resources required to carry out the high-throughput experimentation. This work is performed in collaboration with A. Gilad Kusne, V. Stanev, H. Yu, and A. Mehta. This work is funded by SRC, ONR and AFOSR.

References

[1] Martin L. Green, Ichiro Takeuchi, and Jason R. Hattrick-Simpers, J. Appl. Phys. 113, 231101 (2013).

[2] A. Gilad Kusne, Heshan Yu, Changming Wu, Huairuo Zhang, Jason Hattrick-Simpers, Brian DeCost, Suchismita Sarker, Corey Oses, Cormac Toher, Stefano Curtarolo, Albert V. Davydov, Ritesh Agarwal, Leonid A. Bendersky, Mo Li, Apurva Mehta, Ichiro Takeuchi, Nature Communications 11, 5966 (2020).