Physics-Based Factorized Machine Learning for Predicting Ionic Dielectric Tensors

This repository contains the data and source codes used in the following paper:

Physics-Based Factorized Machine Learning for Predicting Ionic Dielectric Tensors Atsushi Takigawa, Shin Kiyohara, and Yu Kumagai
Phys. Rev. X 16, 021006 (2026) https://doi.org/10.1103/28wr-w896

---

Contents

.
├── common/             # Utility functions and shared modules
├── database/           # Training datasets and screening datasets
├── model/              # Machine learning models and training routines
├── phonon/             # Phonon calculations and feature generation
├── plot_map_eg_vs_totd.py  # Script for plotting the Eg (DDH) vs total dielectric constant (PBEsol) map
├── run_cv.py           # Script for k-fold cross-validation
├── run_fulltrain.py    # Script for training using the full dataset
├── results_screening   # First-principles calculation results for screened candidate materials
└── README.md

---

Dataset

The database directory contains:

• Training dataset: Data for 928 oxides used to train the machine learning model.

The dataset was originally constructed in our previous work [1,2].

• Screening dataset: Data for 8,717 oxides collected from the Materials Project [3] database for large-scale screening.

Training Dataset

• st_pbesol: Crystal structures optimized using the PBEsol exchange–correlation functional.
• dielectric_pbesol: Electronic, ionic, and total dielectric tensors calculated using density functional perturbation theory (DFPT) with the PBEsol functional.
• phonon_pbesol: Phonon eigenfrequencies calculated using DFPT with PBEsol.
• bandgap_ddh: Band gap values calculated using the dielectric-dependent hybrid (DDH) method.

Screening Dataset

• st_mp: Crystal structures collected from the Materials Project database.
• bandgap_mp: Band gap values obtained from the Materials Project database.

---

Screening Results

The results_screening directory contains first-principles calculation results for 142 candidate oxides identified through the machine-learning-based screening.

• st_pbesol: Crystal structures optimized using the PBEsol exchange–correlation functional.
• dielectric_pbesol: Electronic, ionic, and total dielectric tensors calculated using DFPT with PBEsol.
• phonon_pbesol: Phonon eigenfrequencies calculated using DFPT with PBEsol.
• bandgap_ddh: Band gap values calculated using the DDH method.

---

References

[1] Y. Kumagai, N. Tsunoda, A. Takahashi, and F. Oba
Insights into oxygen vacancies from high-throughput first-principles calculations
Phys. Rev. Materials 5, 123803 (2021).
doi:10.1103/PhysRevMaterials.5.123803

[2] Y. Kumagai
Computational screening of p-type transparent conducting oxides using the optical absorption spectra and oxygen-vacancy formation energies
Phys. Rev. Applied 19, 034063 (2023).
doi:10.1103/PhysRevApplied.19.034063

[3] A. Jain, S.P. Ong, G. Hautier, W. Chen, W.D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K.A. Persson
The Materials Project: A materials genome approach to accelerating materials innovation
APL Materials 1, 011002 (2013).
doi:10.1063/1.4812323

---

If you encounter any issues, please contact the authors through the issue tracker or email yukumagai@tohoku.ac.jp.