Prospective on methods of design of experiments for limited data scenarios in materials design and engineering

Machine Learning (ML) is transforming materials discovery and material systems development. Generally, ML methods are best suited to projects with a large amount of readily available data, either within existing databases or that can be gathered via high-throughput experimentation. This is because most ML methods are developed to efficiently extract valuable and often hidden information from large datasets to transform data into knowledge. However, for a significant number of projects in novel materials development, high-throughput experimentation is not an option, rendering databases rather sparse. This occurs from materials discovery through deployment across myriad target applications. The challenge in such projects is not primarily how best to analyze the gathered data but, more importantly, how best to gather data to maximize efficiency in knowledge generation. This paper provides an overview of four basic methods of systematic gathering of experimental data, namely, classical Design of Experiments, Optimal Design, Response Surface Methodology (RSM), and Bayesian Optimization. An introduction and reviews of applications in materials design are provided for each method to demonstrate where and why these methods are used, particularly for materials design and development under limited data scenarios. Graphical abstract