General Graph Neural Network-Based Model To Accurately Predict Cocrystal Density and Insight from Data Quality and...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Cocrystal engineering is crucial for tailoring solid-state properties, with cocrystal density being a key functional parameter.
• Accurate prediction of cocrystal density is essential for material design and discovery.
• Existing methods may lack the precision required for complex cocrystal systems.

Purpose of the Study:

• To develop a robust deep learning framework for accurate cocrystal density prediction.
• To investigate the impact of data quality, feature representation, and model architecture on prediction performance.
• To provide a reliable computational tool for guiding experimental cocrystal design.

Main Methods:

• Development of a graph neural network (GNN)-based deep learning framework.
• Consideration of data quality, feature representation (molecular graphs), and model architecture.
• Inclusion of global attention mechanisms for enhanced feature learning and interpretability.
• Dataset curation focusing on 1:1 stoichiometry cocrystals with data augmentation.

Main Results:

• The GNN model significantly outperforms existing methods in predicting cocrystal density.
• Data quality, particularly stoichiometric ratios, critically influences prediction accuracy.
• Feature representation using molecular graphs with global attention proved effective.
• The model demonstrates high accuracy and generality for predicting densities of unseen cocrystals.

Conclusions:

• The developed GNN framework offers a powerful and generalizable tool for cocrystal density prediction.
• The study provides valuable insights into optimizing machine learning applications in cocrystal engineering.
• This work facilitates experimental investigations and accelerates the discovery of novel cocrystal materials.