Anomaly-Detection-Driven Screening of Thermodynamic Stability from Composition Descriptors Alone.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning in Materials

Background:

• Materials informatics often overlooks novel compositions not present in existing databases.
• Discovering new materials requires efficient methods to explore vast compositional spaces.

Purpose of the Study:

• To develop a composition-based anomaly detection model for predicting material synthesizability.
• To identify key features influencing the stability and potential formation of new materials.

Main Methods:

• An autoencoder model was trained on stable and nearly stable inorganic compounds from the Materials Project database.
• Reconstruction error (Root Mean Square Error - RMSE) was used as an anomaly score.
• Feature importance analysis was performed on element pairs and their electronic configurations (spdf product).

Main Results:

• The model's RMSE correlated with thermodynamic instability (energy above hull).
• Departures from charge neutrality in dummy oxides increased RMSE, even without explicit charge information.
• Element pairs, particularly their spdf product, were key predictors of RMSE, with some pairs like Tantalum (Ta) showing consistent deviations.

Conclusions:

• Composition-only autoencoders can predict material synthesizability and identify potential instabilities.
• Element pair properties and electronic configurations significantly influence material formation.
• The model provides a framework for exploring unregistered compositions in materials informatics.