Predicting materials properties without crystal structure: deep representation learning from stoichiometry.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) accelerates materials discovery by predicting properties efficiently.
• Current ML methods face challenges with input representation: structure-dependent descriptors require known crystal structures, while structure-agnostic methods use limited, hand-engineered features.
• A gap exists for ML models that predict properties using only elemental composition (stoichiometry).

Purpose of the Study:

• To develop a novel machine learning approach for materials discovery.
• To overcome the limitations of current descriptor generation methods in materials science.
• To enable accurate property prediction using only stoichiometric information.

Main Methods:

• Developed a machine learning approach that accepts only the stoichiometric formula as input.
• Treated the stoichiometric formula as a dense weighted graph between elements.
• Implemented an automated descriptor learning process from data.

Main Results:

• The proposed method automatically learns relevant and improvable descriptors from data.
• Achieved lower prediction errors compared to state-of-the-art structure-agnostic methods.
• Demonstrated superior performance with less training data.

Conclusions:

• This graph-based, stoichiometry-input ML approach advances materials discovery.
• The method offers a more efficient and broadly applicable alternative to existing techniques.
• Automated descriptor learning from stoichiometry significantly improves prediction accuracy.