Toward a Generalizable Machine-Learned Potential for Metal-Organic Frameworks

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine-learned potentials (MLPs) offer "quantum-accurate" molecular simulations with linear time complexity, surpassing empirical force fields.
• Generating training data for MLPs from ab initio calculations remains computationally expensive.
• Developing general MLPs for diverse nanoporous metal-organic frameworks (MOFs) is an underexplored area.

Purpose of the Study:

• To investigate the feasibility of creating a single, general MLP applicable to a wide range of Zn-based MOFs with varying chemical and geometric characteristics.
• To reduce the computational burden associated with high-throughput screening of MOFs.

Main Methods:

• Leveraged data-efficient equivariant MLPs.
• Curated a training dataset using density functional theory (DFT) optimized MOF structures.
• Validated the MLP's accuracy in predicting forces and energies on a chemically diverse test set.

Main Results:

• Successfully developed a general MLP for nearly 3000 Zn-based MOFs.
• The MLP reliably predicts physical properties (vibrational, thermodynamic, mechanical) for a large MOF sample.
• Achieved a significant reduction in computational cost for high-throughput screening, enabling investigation of previously inaccessible Zn-MOFs.

Conclusions:

• A general MLP can be effectively developed for a large and diverse set of Zn-based MOFs.
• This approach drastically cuts computational costs, facilitating accelerated discovery of novel MOFs.
• The developed dataset and codes are publicly available to aid future research in complex chemical structures.