Highly Accurate and Fast Prediction of MOF Free Energy via Machine Learning

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Synthesizing novel metal-organic frameworks (MOFs) computationally is hindered by challenges in predicting synthetic accessibility.
• Traditional methods for calculating MOF free energies are computationally expensive, limiting exploration of the vast MOF design space.

Purpose of the Study:

• To develop an efficient machine learning approach for predicting MOF synthetic accessibility and free energy.
• To overcome the computational limitations of traditional methods in assessing MOF synthesis.

Main Methods:

• Curated a dataset of nearly 1 million MOFs (MOFMinE) with calculated properties.
• Developed a novel sequence representation for MOFs (MOFSeq) capturing local and global features.
• Utilized a large language model (LLM-Prop) pretrained on strain energy and fine-tuned for free energy prediction.

Main Results:

• Achieved a mean absolute error of 0.789 kJ/mol_MOFatom for free energy prediction.
• Predicted MOF synthesizability with a 97% F1 score without retraining.
• Correctly identified the lowest free energy MOF polymorph with 78.1% average accuracy.

Conclusions:

• The developed machine learning model efficiently predicts MOF free energy and synthesizability.
• This approach significantly accelerates the identification of viable MOF candidates for experimental synthesis.
• The model demonstrates versatility and high accuracy across various prediction tasks.