Interpretable Graph Transformer Network for Predicting Adsorption Isotherms of Metal-Organic Frameworks.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting adsorbate interactions with metal-organic frameworks (MOFs) is crucial for designing advanced porous materials.
• Existing models struggle with accurate adsorption isotherm prediction, limiting MOF performance optimization.
• Developing robust computational tools is essential for accelerating the discovery of novel MOFs.

Purpose of the Study:

• To introduce MOFNet, a novel deep learning approach for predicting MOF adsorption isotherms.
• To enhance the accuracy and applicability of computational models for gas uptake in MOFs.
• To enable the design of high-performance MOFs through structure-based property prediction.

Main Methods:

• Utilizing a hierarchical representation to encode MOF crystal structures.
• Employing a graph transformer network to capture atomic-level chemical features.
• Implementing a pressure adaptive mechanism with transfer learning for wider pressure range prediction.

Main Results:

• MOFNet significantly outperformed traditional machine learning and graph neural network models on benchmark datasets.
• The model demonstrated high accuracy on real-world experimental adsorption isotherm data.
• The deep learning approach successfully predicted adsorption isotherms for disordered MOFs and imputed missing data.

Conclusions:

• MOFNet offers a powerful and versatile tool for predicting MOF adsorption isotherms.
• The model's ability to interpret structure-property relationships aids in rational MOF design.
• This deep learning approach advances the field of porous materials discovery and optimization.