Modeling ferroelectric phase transitions with graph convolutional neural networks

Acta Physica Sinica, Год журнала: 2024, Номер 73(8), С. 086301 - 086301

Опубликована: Янв. 1, 2024

Ferroelectric materials are widely used in functional devices, however, it has been a long-standing issue to achieve convenient and accurate theoretical modeling of them. Herein, noval approach ferroelectric is proposed by using graph convolutional neural networks (GCNs). In this approach, the potential energy surface described GCNs, which then serves as calculator conduct large-scale molecular dynamics simulations. Given atomic positions, well-trained GCN model can provide predictions forces, with an accuracy reaching up 1 meV per atom. The GCNs comparable that <i>ab inito</i> calculations, while computing speed faster than calculations few orders. Benefiting from high fast prediction model, we further combine simulations investigate two representative materials—bulk GeTe CsSnI₃, successfully produce their temperature-dependent structural phase transitions, good agreement experimental observations. For GeTe, observe unusual negative thermal expansion around region its transition, reported previous experiments. correctly obtain octahedron tilting patterns associated transition sequence. These results demonstrate reliability surfaces for materials, thus providing universal investigating them theoretically.

Язык: Английский

Progress in Protein Pre-training Models Integrated with Structural Knowledge

Acta Physica Sinica, Год журнала: 2024, Номер 73(18), С. 188701 - 188701

Опубликована: Янв. 1, 2024

The AI revolution, sparked by natural language and image processing, has brought new ideas research paradigms to the field of protein computing. One significant advancement is development pre-training models through self-supervised learning from massive sequences. These pre-trained encode various information about sequences, evolution, structures, even functions, which can be easily transferred downstream tasks demonstrate robust generalization capabilities. Recently, researchers have further developed multimodal that integrate more diverse types data. recent studies in this direction are summarized reviewed following aspects paper. Firstly, structures into reviewed: particularly important, for structure primary determinant its function. Secondly, dynamic introduced. may benefit such as protein-protein interactions, soft docking ligands, interactions involving allosteric proteins intrinsic disordered proteins. Thirdly, knowledge gene ontology described. Fourthly, we briefly introduce RNA fields. Finally, most developments designs discuss relationship these with aforementioned information.

Язык: Английский