Speaker
Description
Large-scale atomistic simulations play an important role in the multiscale materials modeling approach. Traditionally, these simulations rely on the classical interatomic potentials. Classical potentials have proven useful for multiple applications thanks to their performance and linear scaling with the system size. However, these potentials often describe chemical behavior of the materials with insufficient accuracy and have limited transferability. Machine-learning (ML) interatomic potentials, a rapidly developing field in the last years, promise to overcome these limitations and while staying computationally efficient ML potentials also can reach the accuracy of the DFT simulations. However, most of the developed ML potentials suffer from the limited transferability as well as the classical potentials and lose their predictive power outside the certain range on volumes, temperatures, compositions, etc.
Recently developed Atomic Cluster Expansion (ACE) ML interatomic potential [1] offers a systematic approach to overcome such issues. Here, we demonstrate the performance of the ACE potential on the example of the transition metals. We implement an efficient and automated parameterization algorithm in order to train the ACE potential on the reference DFT data. We utilize trained potentials to compute a broad spectrum of ground state properties (energy-volume curves, vacancy formation and migration, surface energies, stacking faults, etc.) as well as temperature dependent properties (thermal expansion, melting point, etc.) and show that ACE accurately reproduces the DFT predictions. We also show that ACE accurately describes not only the properties of the bulk material but clusters as well.
[1] “Atomic cluster expansion for accurate and transferable interatomic potentials”, Phys. Rev. B 99, 014104
| Speaker Country | Germany |
|---|