2022 Annual Meeting

Data-Driven Optimization of Ybco Synthesis

Targeted solid-state synthesis plays a major role in driving materials development. However, synthesizing novel compounds with high purity is challenging. Reaction products generally vary with changes in precursors, synthesis temperature, synthesis time, atmosphere, and quality of mixing. Exploring all possible combinations of these variables (i.e., by brute force) calls for many experiments, which is highly inefficient considering the significant costs of experiments. Here, we develop a more intelligent approach to accelerate materials synthesis. With the rise of automation, both in terms of robotics and artificial intelligence, autonomous synthesis-by-design may be achievable such that the yield of arbitrary materials can be optimized with respect to synthesis variables.

Our decision-making algorithm (ARROWS) balances exploration and exploitation by creating an initial ranking by maximizing reaction energy and the number of pairwise interfaces. It updates this ranking based on experimental results by learning possible pairwise reactions and their respective temperatures.

To design and test ARROWS, we present, for the first time, high-throughput synthesis data for YBa2Cu3O7 mapping 188 possible combinations of the chosen precursors and temperatures. Here, we show that our thermodynamically informed algorithm performs better than other state-of-the-art Bayesian Optimization Methods. We further test our algorithm on LiTiPO5 where we optimize for metastable polymorph selectivity and Na2Te3Mo3O16, where we optimize for metastable selectivity against decomposition. In the latter, we vary the synthesis times instead of temperatures to enhance the scope of our method.

We hope that through ARROWS, we can perform experiments sequentially and intelligently, and avoid doing any unnecessary or repetitive experiments.