FFLAME: a fragment-to-framework learning approach for MOF potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Metal-organic frameworks (MOFs) offer vast potential in energy and separation applications due to their structural diversity.
• Accurate property prediction for MOFs is challenging due to their structural flexibility and the limitations of current machine learning potentials (MLPs) which often lack transferability.
• Existing MLPs are typically system-specific, hindering their application across the wide range of MOF structures.

Purpose of the Study:

• To introduce FFLAME (Fragment-to-Framework Learning Approach for MOF Potentials), a novel fragment-centric strategy for training transferable MLPs.
• To enable efficient reuse of chemical environments by decomposing MOFs into metal clusters and organic linkers.
• To reduce the reliance on extensive full-framework training data for developing accurate MOF property prediction models.

Main Methods:

• Developed FFLAME, a fragment-centric machine learning approach for MOF property prediction.
• Decomposed MOFs into constituent metal clusters and organic linkers to facilitate transferable learning.
• Trained MLPs using fragment-informed strategies to improve generalizability and reduce data requirements.

Main Results:

• Fragment-informed training significantly enhances MLP generalizability, especially in data-scarce scenarios.
• FFLAME accelerates model convergence during fine-tuning on new MOFs.
• The approach achieves high accuracy on unseen MOFs with minimal additional training data.

Conclusions:

• FFLAME establishes a robust and data-efficient pathway for developing general-purpose MLPs for diverse framework materials.
• This fragment-based strategy overcomes the limitations of system-specific models, paving the way for broader MLP applicability in MOF research.
• The method promises to accelerate simulations and discovery in materials science by enabling accurate and scalable property predictions.