Data-driven prediction of chemically relevant compositions in multi-component systems using tensor embeddings.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Discovering novel materials is essential for technological advancement.
• Predicting complex multi-component oxide compositions remains a challenge.
• Existing methods often struggle with the vast compositional space of oxides.

Purpose of the Study:

• To develop a predictive model for forecasting complex oxide compositions.
• To leverage data from simpler pseudo-binary oxides for predicting more complex ones.
• To apply tensor decomposition and machine learning for materials prediction.

Main Methods:

• Utilized tensor decomposition (Tucker decomposition) to represent pseudo-binary oxide compositions from the Inorganic Crystal Structure Database (ICSD).
• Extracted tensor embeddings capturing chemical trends like oxidation states and periodic positions.
• Trained a Random Forest classifier using these embeddings to predict the existence of complex oxides.

Main Results:

• The model successfully predicted the existence probabilities of pseudo-ternary and quaternary oxides.
• Achieved high prediction scores, identifying 84% of pseudo-ternary and 52% of quaternary ICSD-registered compositions.
• Demonstrated the efficacy of using simpler oxide data to predict complex material structures.

Conclusions:

• The developed model offers a powerful tool for accelerating the discovery of new functional materials.
• This data-driven approach, using tensor representations, can be extended to other material systems like sulfides and nitrides.
• Highlights the potential of machine learning in navigating complex materials discovery landscapes.