Abstract
The hydrogen storage in MOF-5 and lithiated MOF-5 (Li-MOF-5) is simulated with ring polymer molecular dynamics (RPMD) on machine-learned potentials (MLPs). The MLPs were parametrized based on the message-passing atomic cluster expansion (MACE) foundation model. It was fine-tuned for H2 storage in MOF-5 and Li-MOF-5, using density-functional theory (DFT) single point calculations of each 100th frame from MD trajectories calculated with it. In order to focus the samplings on the most relevant parts of configuration space and reduce the overhead due to erroneous samplings with the foundation model, the fine-tuning was done self-consistently: MD samplings were performed with the fine-tuned MLP, DFT single points calculated and the data used for another fine-tuning, until the property of interest (the spatial H$_2$ distribution in this case) has converged. With this, highly-accurate MLPs were obtained that allowed for the simulation of H2 storage, depending on the temperature and the H2 loading. Nuclear quantum effects, important for the description of H2 at the low temperatures relevant for storage, were accounted for with RPMD. In MOF-5, the H2 molecules are loosely associated to the edges of the framework, whereas in Li-MOF-5, each Li atom coordinates 2-3 H2 molecules, up to room temperature. Its H2 capacity is therefore increased significantly. Nuclear quantum effects lead to a slight decrease in storage capacity for both MOF-5 and Li-MOF-5, in agreement with previous studies. The outlined fine-tuning strategy facilitates the fast and systematic generation of high-quality MLPs for hydrogen storage in arbitrary MOFs.
Supplementary materials
Title
Supporting Information
Description
Input files for the fine-tuning of the MACE foundation model, the sampling of MD trajectories with it and for RPMD simulations with Caracal are given in the Supporting Information.
Actions