|
|
|
# SevenNet: a pretrained universal machine learning force fields
|
|
|
|
|
|
|
|
|
|
|
|
`Name` <be> Prof. Seungwu Han <br>
|
|
|
|
`Affiliation`<be> Department of Materials Science and Engineering, Seoul National University, South Korea
|
|
|
|
|
|
|
|
## Abstract
|
|
|
|
|
|
|
|
Recently, machine learning force fields (MLFFs) have been gaining attention as these methods can overcome the scalability limits of conventional ab initio approaches. However, the long development periods of MLFFs, particularly in multicomponent and reactive systems, hamper their widespread and immediate use. Here we introduce a pretrained universal MLFF, SevenNet-0, that is based on an equivariant graph neural network. SevenNet-0 has been trained over a large amount of crystal structures, and achieved state-of-the-art performance in the Matbench leaderboard. We first demonstrate the generalization capability of SevenNet-0 with the example of Li transport in liquid electrolytes. Although SevenNet-0 was primarily trained on inorganic crystals, it accurately describes the Coulombic interactions between salts and organic solvents.
|
|
|
|
Next, we discuss a multi-fidelity framework that can efficiently improve the accuracy of the pretrained model. Our approach enables the accurate learning of a high-fidelity potential energy surface with minimal high-fidelity data. We test this framework on the InGaN and argyrodite systems. The computational results indicate that geometric and compositional spaces not covered by the high-fidelity meta-gradient generalized approximation (meta-GGA) database can be effectively inferred from low-fidelity GGA data, thus enhancing accuracy and molecular dynamics stability. We also develop a general-purpose SevenNet-MF-0 that utilizes both GGA and meta-GGA data from the Materials Project, significantly enhancing the model performance for high-accuracy tasks such as predicting energies above hull for crystals in general. Furthermore, we demonstrate that the present multi-fidelity learning approach is more effective than either transfer learning or ∆-learning, and can also be extended to learn higher-fidelity levels up to the coupled-cluster level. |
|
|
\ No newline at end of file |
