April 7 - 11, 2025
Seattle, Washington
Symposium Supporters
2025 MRS Spring Meeting & Exhibit
MT02.08.03

High-Throughput Simulations of Dielectric Properties Using Machine Learning

When and Where

Apr 11, 2025
9:00am - 9:30am
Summit, Level 4, Room 423

Presenter(s)

Co-Author(s)

Ryoji Asahi1,Alexander Kutana1

Nagoya University1

Abstract

Ryoji Asahi1,Alexander Kutana1

Nagoya University1
High dielectric permittivity materials are the key component in capacitive electronic and high electric power density applications and devices. Besides the required high relative dielectric permittivity, the desired properties for such materials include temperature and electric field stability, low dielectric losses, and high breakdown voltage. Most of the presently known materials with high permittivity do not meet all of these conditions, limiting applications. On the other hand, in In-Nb co-doped rutile TiO2, permittivity boosting to >104 was reported, suggesting that it is possible to intrinsically improve the dielectric constant of bulk crystals by controlling the complex defect structure [1,2].
The dielectric properties are evaluated by state-of-the-art density functional perturbation theory, which is rather computationally demanding. In addition, complexity of dielectric properties including doping effects, temperature effects, and phase transitions makes us difficult to perform systematic or high-throughput evaluation of wide-range of dielectric materials.
In this study, we employ graph convolutional neural network (GNN) model to predict dielectric constants in large and complex material systems. Using this GNN model, the "local dielectric constants" are assigned at each atomic site in the network layer, indicating that the GNN model captures the features of the local structure, for example, modified by doping. To calculate the dependence on temperature and doping concentration, it is necessary to evaluate the statistical average using a large-scale model. To this end, we have developed an algorithm combining Monte Carlo sampling and the GNN model. In particular, the newly developed rotationally equivariant GNN model accurately predicts atomic Born effective charge and dielectric tensors [3]. The proposed method demonstrates accurate evaluation of temperature dependency of dielectric properties.

References
(1) Kutana, Shimano, Asahi, Sci. Rep. 13, 3761 (2023).
(2) Shimano, Kutana, Asahi, Sci. Rep. 13, 22236 (2023).
(3) Kutana, Shimizu, Watanabe, Asahi, arXiv:2409.08940

Symposium Organizers

Ling Chen, Toyota North America
Bin Ouyang, Florida State University
Chris Bartel, University of Minnesota
Eric McCalla, McGill University

Symposium Support

Bronze
GE Vernova's Advanced Research Center

Session Chairs

Chris Bartel
Bin Ouyang

In this Session