MOFClassifier: A Machine Learning Approach for Validating Computation-Ready Metal-Organic Frameworks.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• High-quality structural data is crucial for computational discovery of metal-organic frameworks (MOFs).
• Existing MOF databases contain significant errors, hindering efficient screening.
• Current rule-based error-checking methods have limitations and misclassify structures.

Purpose of the Study:

• To develop a novel machine learning approach for accurate classification of computation-ready MOFs.
• To overcome limitations of existing methods in identifying structural and chemical errors in MOF data.
• To improve the reliability of large-scale computational screening for new MOF materials.

Main Methods:

• Developed MOFClassifier, a machine learning model using a positive-unlabeled crystal graph convolutional neural network (PU-CGCNN).
• The model learns patterns from perfect crystal structures to predict a "crystal-likeness score" (CLscore).
• Evaluated performance using ROC values and compared against existing rule-based methods.

Main Results:

• MOFClassifier achieved an ROC value of 0.979, surpassing the previous best of 0.912.
• The model successfully identified subtle structural and chemical errors missed by current methods.
• Accurately recovered misclassified false-negative structures, reducing the risk of overlooking potential MOF candidates.

Conclusions:

• MOFClassifier offers a significant advancement in accurately classifying MOFs for computational screening.
• The tool enhances the efficiency and reliability of discovering new MOF materials.
• Freely available and integrated into the CoRE MOF DB 2025 v1.0, accelerating MOF research.