Predicting thermodynamic stability of inorganic compounds using ensemble machine learning based on electron...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) accelerates the discovery of new compounds by predicting thermodynamic stability.
• Traditional methods are time-consuming and resource-intensive.
• Existing ML models may have performance limitations due to domain-specific knowledge and inherent biases.

Purpose of the Study:

• To develop a novel ML framework for predicting compound thermodynamic stability.
• To enhance prediction accuracy and efficiency by integrating diverse domain knowledge.
• To demonstrate the framework's utility in exploring novel material compositions.

Main Methods:

• A machine learning framework based on electron configuration was developed.
• Stack generalization was employed, integrating two additional models with diverse domain knowledge.
• The model's performance was validated using experimental results and first-principles calculations.

Main Results:

• The proposed model achieved a high Area Under the Curve score of 0.988 for stability prediction.
• The model demonstrated exceptional sample utilization efficiency, requiring only one-seventh of the data of existing models.
• The framework successfully identified stable compounds in unexplored composition spaces, including two-dimensional wide bandgap semiconductors and double perovskite oxides.

Conclusions:

• The electron configuration-based ML framework accurately predicts compound thermodynamic stability.
• This approach offers significant improvements in efficiency and data utilization compared to existing methods.
• The framework provides a versatile tool for accelerating the discovery of novel materials.