DFT-Quality Adsorption Simulations in Metal-Organic Frameworks Enabled by Machine Learning Potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Metal-organic frameworks (MOFs) show promise for adsorption and separation, with grand canonical Monte Carlo (GCMC) simulations widely used for screening.
• Empirical force fields in GCMC limit accuracy for MOFs with open-metal sites, while density-functional theory (DFT) is computationally prohibitive for routine GCMC.
• A need exists for accurate, computationally feasible methods to predict gas adsorption in MOFs.

Purpose of the Study:

• To develop and validate a protocol for training machine learning potentials (MLPs) using DFT data for accurate GCMC simulations of CO2 adsorption in MOFs.
• To assess the performance of MLPs derived from DFT calculations in predicting adsorption isotherms and heats of adsorption.
• To investigate the transferability of MLPs across different MOF structures.

Main Methods:

• Training MLPs using DFT-calculated intermolecular interaction energies and forces for CO2 in ZIF-8 and Mg-MOF-74.
• Utilizing the equivariant NequIP model for efficient data training.
• Performing GCMC simulations with the trained MLPs to derive adsorption isotherms and heats of adsorption.

Main Results:

• Achieved high accuracy (interaction energy error < 0.2 kJ mol⁻¹ per adsorbate) for CO2 in ZIF-8 using MLPs.
• Obtained accurate adsorption isotherms and heats of adsorption for CO2 in ZIF-8 and Mg-MOF-74.
• Demonstrated that MLPs trained on one MOF (ZIF-8) do not accurately predict adsorption in others (ZIF-3, ZIF-4, ZIF-6) without specific training data, highlighting the need for tailored or general-purpose MLPs.

Conclusions:

• The proposed protocol effectively integrates DFT accuracy with GCMC efficiency via MLPs for studying gas adsorption in nanoporous materials.
• MLP accuracy is highly dependent on the training data, emphasizing the importance of representative datasets for specific MOFs or the development of generalizable MLPs.
• This methodology paves the way for highly accurate, first-principles investigations of guest adsorption in diverse nanoporous materials.