Comparison of Universal Graph Neural Network Force Fields: Molecular Dynamics Calculations for a Lithium-Ion Conductor#
Universal Graph Neural Network (GNN) force fields achieve higher versatility and accuracy than conventional force fields by utilizing neural networks. Many universities, research institutions, and companies have developed and released them. AdvanceSoft has modified LAMMPS to support various GNN force fields, making them available through Advance/NanoLabo.
In this case study, as an example of comparing various GNN force fields, we perform a molecular dynamics calculation on the lithium-ion conductor Li3OCl. We will demonstrate its applicability to multi-element systems and verify whether calculations can be performed at high temperatures without structural collapse.
Model Creation and Calculation Conditions#
Based on the structure file of Li3OCl (mp-985585) obtained from the Materials Project, a 2×2×2 supercell model was created. One Li atom was removed from this model to introduce a vacancy, resulting in a 39-atom model.
The molecular dynamics calculation was performed with an NVT ensemble at 1500K, using a timestep of 2 fs. By enabling the "Diffusion Coefficient" feature in Advance/NanoLabo, quantities related to atomic diffusion are calculated and plotted.
Results#
We performed molecular dynamics calculations using each GNN force field and plotted the mean square displacement of Li atoms over time.
All force fields were able to simulate the conduction of Li atoms without the structure collapsing. Differences in conductivity were observed depending on the force field, and a comparison with first-principles calculations is necessary for a quantitative performance evaluation. Such simulations can be applied to tasks like searching for compositions of lithium-ion conductors with high conductivity.