Recent advances in machine learning methods for materials discovery have made considerable progress in reducing the time to identify novel materials with tailored properties. Materials domains such as high entropy materials, polymers, semiconductors, and molecules have seen numerous materials-by-design algorithmic advances over the last decade, showing the power of machine learning in aiding the exploration of large chemical and structural spaces. However, a major bottleneck in these methodologies still exists, particularly in their ability to extrapolate to new design spaces. Existing machine learning algorithms are built on a foundation of interpolative mathematical formalisms, allowing them to identify new candidates that are different from known materials but still relatively similar. These algorithms still struggle to extrapolate to completely novel material systems, often referred to as “out-of-domain” samples, as their underlying mathematical formalisms are built upon parameterized similarity metrics fit to existing materials databases. By design, these algorithms are biased toward existing data, and often break down as similarity to known materials breaks down. Recently, out-of-domain machine learning methods have made progress toward alleviating this challenge, often by including some level of known physics in the algorithms. In this article, we explore these recent advances from the perspective of materials-by-design, showcasing how out-of-domain machine learning has made progress in identifying truly novel material candidates, and discuss the remaining challenges.
Tenorio et al. (2026) studied this question.