|
Efficient determination of Born-effective charges, LO-TO splitting, and Raman tensors of solids with a real-space atom-centered deep learning approach
O. Malenfant-Thuot, K. Ryczko, I. Tamblyn, M. Cote Journal of Physics: Condensed Matter, 36, 425901 (2024) Predicting how light interacts with crystalline materials typically requires computationally expensive quantum mechanical simulations, particularly for properties like Raman spectra that depend on subtle electronic response functions. In this work, we introduce RADNET (Real-space Atomic Decomposition NETwork), a deep learning framework that can rapidly and accurately predict polarization and dielectric properties of solids using an atom-centered approach. By leveraging automatic differentiation, RADNET efficiently computes Born-effective charges, LO-TO splitting frequencies, and Raman tensors from learned representations, achieving excellent agreement with traditional ab initio calculations on test materials such as GaAs and BN. This approach makes it possible to predict optical response properties for larger, more complex structures at a fraction of the computational cost of conventional methods. |
