MOFormer: Self-Supervised Transformer Model for Metal-Organic Framework Property Prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Metal-organic frameworks (MOFs) possess high porosity, enabling diverse applications.
• The vast chemical space of MOFs necessitates efficient screening for optimal material discovery.
• Current high-throughput screening methods, like DFT, are computationally intensive and require 3D structures.

Purpose of the Study:

• To propose a novel, structure-agnostic deep learning method for accelerated MOF property prediction.
• To circumvent the need for 3D atomic structures in evaluating hypothetical MOFs.
• To enhance the efficiency and accuracy of MOF screening for specific applications.

Main Methods:

• Developed MOFormer, a Transformer-based deep learning model for MOF property prediction.
• Utilized a structure-agnostic approach, taking MOF text string representations (MOFid) as input.
• Introduced a self-supervised learning framework to pretrain MOFormer using >400k public MOF data, correlating structure-agnostic and structure-based representations.

Main Results:

• MOFormer achieved state-of-the-art structure-agnostic prediction accuracy compared to existing descriptors.
• Pretraining the MOFormer model significantly improved prediction accuracy for downstream tasks.
• MOFormer demonstrated superior data efficiency for quantum-chemical property prediction compared to structure-based models like CGCNN, especially with limited data.

Conclusions:

• MOFormer offers a novel and efficient deep learning perspective for MOF property prediction.
• The structure-agnostic approach accelerates the screening process for discovering new MOFs.
• Self-supervised pretraining enhances the performance and data efficiency of MOF property prediction models.