Coarse-Grained Crystal Graph Neural Networks for Reticular Materials Design

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Reticular materials like metal-organic frameworks (MOFs) and covalent organic frameworks (COFs) have diverse applications.
• Predicting properties of these materials is complex due to vast chemical space.
• Current AI models use atomic-level graphs, which can be computationally expensive and include redundant features.

Purpose of the Study:

• To develop a more efficient materials representation for property prediction.
• To overcome limitations of atomic-level graph neural networks in reticular materials design.
• To introduce a coarse-grained crystal graph approach for inverse materials design.

Main Methods:

• Developed a coarse-grained crystal graph representation focusing on molecular building units.
• Assessed neural network performance using composition-based, crystal-structure-aware, and coarse-grained models.
• Evaluated predictive accuracy and energy efficiency of different representations.

Main Results:

• Coarse-grained crystal graph neural networks demonstrated competitive accuracy with significantly lower computational costs.
• The proposed method is a viable alternative to atomic-level graph neural networks.
• Models were successfully integrated into an inverse materials design pipeline.

Conclusions:

• The coarse-grained crystal graph framework offers an efficient and accurate approach for reticular materials property prediction.
• This method challenges the traditional atom-centric perspective in materials design.
• It provides a valuable tool for accelerating the discovery of novel reticular materials.