Metal: Active Learning Advances High Entropy Oxide Discovery with Machine Learning Interatomic Potentials

The search for novel materials is driving innovation in sustainable technologies, and high entropy oxides are emerging as a promising, yet largely unexplored, area of research. Joakim Brorsson, Henrik Klein Moberg, and Joel Hildingsson, all from the Department of Physics at Chalmers University of Technology, alongside Jonatan Gastaldi, Tobias Mattisson, and Anders Hellman, have demonstrated a powerful new approach to accelerate materials discovery in this field. Their work focuses on identifying optimal oxygen carriers for chemical looping, a process vital for efficient energy production and carbon capture. By combining active learning strategies with first-principles calculations and advanced machine learning interatomic potentials, the researchers have not only validated their methodology on high entropy perovskites, but also successfully identified specific high entropy oxide compositions with exceptional oxygen transfer capabilities. This research highlights the transformative potential of active learning, suggesting it will rapidly become an indispensable tool for materials scientists seeking to navigate complex compositional landscapes.

Machine Learning Accelerates Oxygen Carrier Discovery

The discovery of suitable oxygen carriers has been tentatively explored, but is now becoming more accessible thanks to machine learning. This research tackles the task of finding oxygen carriers for chemical looping processes by leveraging active learning strategies combined with first-principles calculations. High efficiency has been achieved by exploiting recently developed machine learning interatomic potentials, and the proposed approaches were validated using an established computational framework for identifying high entropy perovskites for chemical looping combustion. This research objectively aims to accelerate the discovery of novel oxygen carriers, circumventing the limitations of traditional trial-and-error methods.

The approach integrates active learning with density functional theory calculations to efficiently screen a large compositional space of high entropy perovskites. Machine learning models are trained on calculated formation energies and oxygen vacancy formation energies, guiding the selection of promising candidates and reducing the need for computationally expensive first-principles calculations. A key contribution of this work is a robust and efficient workflow for materials discovery. The combination of active learning and first-principles calculations allows for the rapid identification of high entropy perovskites with enhanced oxygen carrier performance. The developed machine learning interatomic potentials provide an accurate and computationally affordable means of predicting material behaviour, paving the way for the exploration of even more complex compositional spaces and offering a powerful tool for designing next-generation oxygen carriers.

Bayesian Optimisation of Oxygen Carrier Materials The research

Scientists achieved a significant breakthrough in materials discovery by successfully applying active learning strategies combined with first-principles calculations to identify high entropy oxides. The research focused on discovering oxygen carriers crucial for chemical looping processes, a technology with potential for efficient fuel conversion and carbon capture. Experiments revealed that greedy and Thompson-based sampling, guided by uncertainty estimates from Gaussian processes, were the most effective strategies for navigating complex compositional spaces. Building on this initial success, the team applied the refined methodology to the more challenging task of discovering high entropy oxygen carriers specifically for chemical looping oxygen uncoupling.

Results demonstrate the generation of both qualitative and quantitative outcomes, culminating in detailed lists of materials exhibiting high oxygen transfer capacities and configurational entropies. The best performing candidates were found to be based on CaMnO3, incorporating a diverse range of additional species, including titanium, cobalt, copper, and unexpectedly, yttrium and samarium. Measurements confirm that active learning approaches are critical for accelerating materials discovery, effectively reducing the need for exhaustive trial-and-error methods. The study leveraged a Wasserstein auto-encoder neural network trained on compositions and target properties, estimated using a machine learning interatomic potential validated by first-principles data. This streamlined approach allowed scientists to efficiently explore the vast compositional space of high entropy materials, identifying candidates for three distinct chemical looping applications: dry reforming, air separation, and oxygen uncoupling, delivering a powerful new tool for materials scientists.

High Entropy Oxide Discovery via Active Learning Scientists

This work demonstrates the successful application of active learning strategies, combined with first-principles calculations and machine learning interatomic potentials, to the challenging problem of high entropy oxide discovery. Researchers focused on identifying materials suitable for chemical looping processes, specifically oxygen carriers for both oxygen uncoupling and air separation. Through this approach, they identified promising candidates based on CaMnO3 and LaMnO3 perovskites, incorporating elements like titanium, cobalt, copper, yttrium, and strontium, some of which were unexpected. The findings highlight the efficacy of active learning, particularly greedy and Thompson-based sampling, in navigating complex compositional spaces and accelerating materials discovery.

While the study revealed a tendency towards lanthanum and iron-rich compositions, potentially influenced by the initial training data, the use of techniques like Monte Carlo dropout helped broaden the compositional range. Differences were also observed between the optimal conditions for chemical looping oxygen uncoupling and air separation, with greedy sampling proving slightly more effective for the latter. The authors acknowledge that the initial training data may have introduced some bias towards certain elements, and further investigation is needed to fully understand this influence. Future research could focus on expanding the compositional space explored and refining the machine learning models to improve predictive accuracy. Nevertheless, this work establishes a powerful methodology for materials discovery, suggesting that active learning will become increasingly integral to the field.