Material exploration through active learning -- METAL

Abstract

The discovery and design of new materials are paramount in the development of green technologies. High entropy oxides represent one such group that has only been tentatively explored, mainly due to the inherent problem of navigating vast compositional spaces. Thanks to the emergence of machine learning, however, suitable tools are now readily available. Here, the task of finding oxygen carriers for chemical looping processes has been tackled by leveraging active learning-based strategies combined with first-principles calculations. High efficiency and efficacy have, moreover, been achieved by exploiting the power of recently developed machine learning interatomic potentials. Firstly, the proposed approaches were validated based on an established computational framework for identifying high entropy perovskites that can be used in chemical looping air separation and dry reforming. Chief among the insights thus gained was the identification of the best performing strategies, in the form of greedy or Thompson-based sampling based on uncertainty estimates obtained from Gaussian processes. Building on this newfound knowledge, the concept was applied to a more complex problem, namely the discovery of high entropy oxygen carriers for chemical looping oxygen uncoupling. This resulted in both qualitative as well as quantitative outcomes, including lists of specific materials with high oxygen transfer capacities and configurational entropies. Specifically, the best candidates were based on the known oxygen carrier CaMnO3 but also contained a variety of additional species, of which some, e.g., Ti; Co; Cu; and Ti, were expected while others were not, e.g., Y and Sm. The results suggest that adopting active learning approaches is critical in materials discovery, given that these methods are already shifting research practice and soon will be the norm.