Towards the automatic exploration of materials by machine learning

Accelerated mapping of composition-structure-property relationships is a central goal in current materials research. Advanced approaches leverage machine learning (ML) models on existing databases, including data mining from the literature. However, efficient and effective application of these ML models requires sufficiently large and consistent experimental datasets, ideally measured under identical conditions. Therefore, combinatorial materials synthesis—parallelized methods that create libraries of materials with systematic parameter variations—combined with high-throughput characterization capabilities, is crucial for AI-driven materials discovery.Among various material families, perovskite oxides (ABO3) offer a vast compositional space with wide applications due to their structural and chemical flexibility. In a recent paper published in Advanced Materials, we presented a comprehensive methodology for studying entire families of perovskite oxides for energy applications [1]. Specifically, we investigated the composition-performance relationships of high-entropy La0.8Sr0.2MnxCoyFezO3±𝞭 perovskite oxides (0 < x, y, z < 1; x + y + z ≈ 1) as oxygen electrodes in Solid Oxide Cells. By depositing a continuous compositional map using thin-film combinatorial pulsed laser deposition, we obtained experimental data on structural, compositional, and functional properties for the entire material family through six advanced characterization methodologies with mapping capabilities. We demonstrated that supervised machine learning methods, particularly random forests, effectively capture the complex relationships between composition, structural features, and electrochemical performance, including oxygen transport properties. Using these predictive methods, we created an accurate continuous performance map for the entire compositional space under study and made it available to the community through an open database [2].