Fast and Accurate Machine Learning-enabled Energy Models and Atomistic Simulations of Chemically-Complex Nanomaterials
Hsin-An Chen1, Ping-Han Tang2, Svetozar Najman3, Po-Yu Yang3, Chun-Wei Pao3*
1Department of Materials and Resource Engineering, National Taipei University of Technology, Taipei, Taiwan
2Molecular Science and Digital Innovation Center, GGA Corporation, Taipei, Taiwan
3Research Center for Applied Sciences, Academia Sinica, Taipei, Taiwan
* Presenter:Chun-Wei Pao,
The chemically complex nanomaterials are drawing increasing attentions in the materials community. The structures of these nanomaterials, for example, chemical short-range orders or phase segregations play important roles for their properties; however, structural characterization of these complex materials at nanoscale poses grand challenges. From theorists' perspective, exploration of possible nanostructures is computationally prohibitive for first-principle calculations due to both system size and computation time limitations. Herein, we demonstrate that by utilizing machine-learning-enabled energy predictors trained from thousands or tens of thousands atomistic images labeled with energies/atomic forces from DFT calculations, it is possible to perform extensive sampling of nanostructures of chemically complex materials such as mixed ion perovskite, 2D perovskite, or even high entropy materials with high fidelity to respective DFT calculations, thereby providing nanoscale structural insights into these chemically complex materials, which are extremely difficult to be extracted from both DFT calculations and current state-of-the-art experimental characterization techniques.

Keywords: chemically complex materials, nanostructure, atomistic simulations, machine learning