Mlatticeabc: Generic Lattice Constant Prediction of Crystal Materials Using Machine Learning

Area of Science:

• Materials Science
• Computational Materials Science
• Crystallography

Background:

• Lattice constants (unit cell edge lengths and plane angles) are crucial for understanding crystal structures and predicting material properties.
• Previous machine learning models achieved moderate success (R²=0.82 for cubic crystals) but struggled with diverse material compositions.
• Specialized models for fixed material families (e.g., perovskites) showed higher performance but lacked general applicability.

Purpose of the Study:

• To develop a machine learning model for accurate and generalizable prediction of crystal lattice constants.
• To introduce a novel descriptor set for enhanced lattice parameter prediction.
• To improve the prediction of lattice angles for various crystal systems.

Main Methods:

• Development of MLatticeABC, a random forest machine learning model.
• Utilization of a new set of descriptors based on molecular formulas.
• Training and validation on a diverse dataset of crystal materials.

Main Results:

• Achieved an R² score of 0.973 for lattice parameter 'a' in cubic crystals.
• Obtained an average R² score of 0.80 for lattice parameter 'a' across all crystal systems.
• Demonstrated significant performance improvements for lattice angle predictions (R² between 0.498 and 0.757 for b and c).

Conclusions:

• MLatticeABC offers a robust method for predicting lattice constants from molecular formulas.
• The model shows broad applicability to diverse crystal materials, overcoming limitations of previous approaches.
• The developed model and source code are publicly available to facilitate further research in materials science.