Materials informatics (MI) is attracting attention as a new approach in materials science, and progress has been made in building predictive models using machine learning while utilizing first-principles calculations.
Materials informatics (MI) is attracting attention as a new approach in materials science, and progress has been made in building predictive models using machine learning while utilizing first-principles calculations and molecular dynamics (MD) calculations. However, there are some major challenges to the practical application of MI. First-principles calculations (DFT) have very high accuracy, but are extremely computationally expensive and are not suitable for large-scale materials exploration. Additionally, molecular dynamics (MD) calculations require an appropriate force field, which is often difficult to apply to new material systems. Neural network potential (NNP), which uses machine learning, is attracting attention as a solution. In this article,The purpose is to evaluate how accurate and practical M3GNET (Materials Graph Neural Network) can be used.Let's say. In particular, we will clarify the advantages and limitations of M3GNET by comparing it with the first-principles calculation software Quantum ESPRESSO (QE), and explore in what situations it can be used.
In this study, in order to evaluate the adsorption behavior of water (H₂O) on the surface of platinum (Pt) slabs, we analyzed changes in adsorption energy and interatomic distance, and compared the calculation accuracy and calculation time of QE and M3GNET.
In this study, we compared the accuracy of M3GNET and QE by evaluating the adsorption energy of water on Pt slabs. Specifically, we calculated the energy values of a Pt slab, a simple H₂O, and a Pt-H₂O model in which H₂O is placed on a Pt slab, and found the difference between them. The table below compares the energy values (eV) in QE and M3GNET. ΔE is calculated using formula (1).
ΔE = E (Pt-H₂O) - E (Slab Pt) - E (H₂O) (1)
These results show that although M3GNET can reproduce qualitative trends, there are slight differences in the accuracy of energy values. Although there were small differences in ΔE, their signs were consistent. When comparing the optimization results of the Pt-OH-H model in which H₂O is placed on a Pt slab, it was found that the position of Pt attached after OH and H separation is different, so the appearance of the structure differs depending on the viewpoint, but the optimized structure itself is similar, and the optimized structure of M3GNET was in good agreement with QE.（Figure 2）.
There was a noticeable difference in calculation time. The figure below shows a comparison of calculation time (Wall Time) between QE and M3GNET.
As is clear from this graph, it was confirmed that M3GNET can perform energy calculations in a significantly shorter time than QE.
In this way, M3GNET achieves several hundred times faster calculation speed than first-principles calculations, and is considered to be extremely useful in large-scale materials exploration and molecular dynamics (MD) simulations.
In this article, we compared M3GNET with DFT in terms of accuracy and calculation time. As a result, M3GNET is extremely useful in large-scale materials exploration.
Although there were some differences in ΔE, the M3GNET results were qualitatively consistent with QE, confirming that the optimized structures were also similar. In particular, M3GNET was overwhelmingly faster in terms of calculation time, with the Pt model being approximately 356 times faster, the H₂O model being approximately 23 times faster, and the Pt-OH-H model being approximately 610 times faster.DFT is still required for detailed physical property analysis such as electronic structure and band calculationsHowever,M3GNET is very useful for large-scale molecular dynamics simulations and screeningThis has been confirmed. By using M3GNET, it becomes possible to perform large-scale simulations that were previously impossible due to time and computational resources.
Furthermore, M3GNET is also useful in MI and can be a powerful tool for rapidly screening material candidates and predicting their properties. It has the potential to accelerate the discovery and design of new materials by efficiently proceeding with large-scale materials searches that take time with conventional first-principles calculations.
In the next article, we will introduce specific ways to use M3GNET for screening and material prediction.
keywords:
#M3GNET #GNN #QuantumESPRESSO #LAMMPS #MaterialsInformatics #MI
The content is prepared by: kazuki.mori@mat3ra.com
