Machine learning for combinatorial exploration of quantum materials

We have been applying machine learning (ML) to high-throughput experimentation in a variety of ways in order to discover new quantum materials [1]. The main focus of our exploration have been superconductors and topological insulators.  The combinatorial strategy can be used to uncover minute details of composition - structure – property relationships which reflect the physical origin of superconductivity [2]. ML can substantially enrich combinatorial experimentation in guiding the experiments as well as streamlining the massive amount of data which result from combinatorial libraries. For instance, we have developed ML models for superconducting critical temperature which can be used to predict possible new superconductors. Our recent emphasis is to develop autonomous combinatorial experimentation techniques based on Bayesian active learning which can further speed up the screening process by removing the necessity to test every composition on libraries. We have shown that we can reduce the number of experiments by an order of magnitude in this manner. I will discuss our recent live demonstration where autonomous interaction between experiment and theory is used to rapidly map thin-film phase diagrams of multinary systems in closed-loop cycles. This work is carried out in collaboration with A. G. Kusne, V. Stanev, H. Liang, H. Yu, J. Park, and J. Paglione.