MOFSimBench: evaluating universal machine learning interatomic potentials in metal-organic framework molecular...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Universal machine learning interatomic potentials (uMLIPs) offer accurate and efficient atomistic simulations.
• Metal-organic frameworks (MOFs) are crucial for carbon capture, energy storage, and catalysis.
• Modeling MOFs presents challenges for uMLIPs due to their complexity and lack of training data.

Purpose of the Study:

• To introduce MOFSimBench, a benchmark for evaluating uMLIPs on nanoporous materials.
• To assess the performance of various uMLIP architectures on key modeling tasks.
• To guide the adoption and development of uMLIPs for MOF applications.

Main Methods:

• Developed MOFSimBench for evaluating uMLIPs on structural optimization, MD stability, bulk properties, and host-guest interactions.
• Evaluated 20 uMLIP models of diverse architectures.
• Compared uMLIP performance against classical force fields and fine-tuned ML potentials.

Main Results:

• Top-performing uMLIPs consistently outperformed classical force fields and fine-tuned ML potentials.
• uMLIPs demonstrated readiness for practical nanoporous materials modeling.
• Data quality was identified as more critical than model architecture for uMLIP performance.

Conclusions:

• MOFSimBench provides a valuable resource for assessing and advancing uMLIPs for nanoporous materials.
• The study validates the potential of uMLIPs for real-world applications in MOF modeling.
• Future uMLIP development should prioritize high-quality training data.