Connecting metal-organic framework synthesis to applications using multimodal machine learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Hundreds of thousands of new materials, including over 5000 metal-organic frameworks (MOFs) annually, are synthesized with unique properties.
• Identifying optimal applications for newly synthesized materials is a significant challenge, hindering their broader use.
• Existing methods often require extensive characterization, delaying application discovery.

Purpose of the Study:

• To develop a multimodal approach for predicting material properties and applications from synthesis data.
• To create a synthesis-to-application map for MOFs to guide material selection.
• To identify novel applications for existing and newly synthesized MOFs.

Main Methods:

• Utilized powder X-ray diffraction (PXRD) patterns and synthesis chemicals as input data.
• Employed self-supervised pretraining on crystal structures from MOF databases.
• Assessed model robustness against experimental imperfections and augmented with a recommendation system.

Main Results:

• Achieved accurate predictions for diverse properties (pore structure, chemical, quantum-chemical) even with limited data.
• Demonstrated the method's robustness against experimental measurement imperfections.
• Generated a synthesis-to-application map, revealing optimal material classes for various uses.

Conclusions:

• The multimodal approach effectively predicts MOF properties and applications using readily available synthesis information.
• The developed tool, including open-source code and a web app, accelerates the discovery of industrial applications for new materials.
• This work facilitates the efficient matching of novel materials with their most promising uses, driving innovation.