About this Abstract |
| Meeting |
MS&T25: Materials Science & Technology
|
| Symposium
|
Grain Boundaries, Interfaces, and Surfaces: Fundamental Structure-Property-Performance Relationships
|
| Presentation Title |
Neural-Network Potential Based on Trainable Descriptor for Modeling Complex Interfacial Structures and Properties |
| Author(s) |
Masami Uchida, Tatsuya Yokoi, Yu Ogura, Katsuyuki Matsunaga |
| On-Site Speaker (Planned) |
Masami Uchida |
| Abstract Scope |
Artificial neural network (ANN) potentials trained with DFT data are promising for exploring interfacial atomic structures. However, their predictive power is often limited when diverse atomic environments are involved, due to the use of fixed analytic structural descriptors. In this study, we propose a trainable descriptor in which the analytic functions are replaced by two- and three-body ANNs, referred to as the ANN descriptor. All ANNs, including the descriptor, are trained simultaneously on given datasets, enabling numerical optimization of the entire functional form. To demonstrate the effectiveness of this approach, we investigate asymmetric tilt grain boundaries in silicon. The ANN descriptor successfully captures complex atomic arrangements and reveals unique atomic and electronic structures at the interfaces. This highlights the potential of the ANN descriptor for advancing fundamental insights into the structure–property relationships of functional interfaces. |