Machine learning in materials design and discovery: Examples from the present and suggestions for the future

We provide a brief discussion of "What is machine learning?" and then give a number of examples of how these methods have recently aided the design and discovery of new materials, such as new shape memory alloys, with enhanced targeted properties, such as lower hysteresis. These examples illustrate how discoveries can be made from large databases, for example, those generated by high throughput DFT calculations and also how they can be made from experimentally growing smaller databases in an active learning manner. Additionally, we discuss such advanced machine learning methods as multiobjective and multifidelity optimization that permit proposing new materials with the simultaneous optimization of more than one targeted property, such as a material with low hysteresis and high Curie temperature, and permit using fewer costly experiments and calculations by combining them with less costly ones to achieve modeling comparable to using only many costly ones. We conclude with a brief discussion of future machine learning opportunities in the context of high throughput experiment and on-the-fly adjustment of synthesis. More speculatively, we end by discussing how might we mesh materials science more fittingly with machine learning.