Prescreening Point Defects in Semiconductors With Machine Learning
Abstract
High-throughput calculations using density-functional theory (DFT) are commonly used to explore point defects for applications in power electronics and quantum technologies. There is currently a major shift away from these traditional simulation techniques towards machine learning (ML) methods. We explore a class of physics-guided ML models for predicting defect formation energies and zero-phonon lines (ZPL) to identify point defects for quantum applications. The models are specifically targeted for use in a prescreening step for accelerated high-throughput workflows, and are therefore designed to avoid the costly relaxation step typically present with ML interatomic potentials (MLIPs). We compare performance for single and double point defect systems in 4H-SiC with ridge, kernel ridge, and multilayer perceptron (MLP) models using three different descriptors representing the defect systems. For vacancies and substitutions, the optimized models give mean absolute errors (MAEs) of 0.437 eV for the formation energy and 0.202 eV for ZPLs, which is just above the level at which such predictions can be useful even beyond the targeted prescreening, i.e., in some applications they may completely replace the need for costly DFT calculations. For interstitials the MAEs are larger, 1.101 eV for the formation energy and 0.230 eV for the ZPL, which, while still useful for prescreening, will not generally be useful for more detailed characterization. Hence, while the results may be further improved by model design and optimization, the models presented in this work are already useful for prescreening in high-throughput characterization of point defects.
Related
- DefectSkills: Guiding AI agents through first-principles calculations for point defects in semiconductors and insulators
- Small-supercell and Small-dataset Training Strategy of Machine Learning Interatomic Potentials for Point Defects
- Sorption Molecular Dynamics of Point Defects in 2D Materials with a Machine Learning Potential
- One-point functions of the Lagrangian in 𝒩=4 SYM with boundaries, defects and interfaces
- Matrix Density Waves and Fractionally Charged Point Defects in Flavor Weyl Semimetals
BibTeX
@misc{karlsson2026prescreening,
title = {{Prescreening Point Defects in Semiconductors With Machine Learning}},
author = {Paul Karlsson and Joel Davidsson and Rickard Armiento},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.14846},
eprint = {2609.14846},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.14846},
}
RIS
TY - UNPB
TI - Prescreening Point Defects in Semiconductors With Machine Learning
AU - Paul Karlsson
AU - Joel Davidsson
AU - Rickard Armiento
PY - 2026
DA - 2026/09/13
PB - arXiv
DO - 10.48550/arxiv.2609.14846
UR - https://arxiv.org/abs/2609.14846
AB - High-throughput calculations using density-functional theory (DFT) are commonly used to explore point defects for applications in power electronics and quantum technologies. There is currently a major shift away from these traditional simulation techniques towards machine learning (ML) methods. We explore a class of physics-guided ML models for predicting defect formation energies and zero-phonon lines (ZPL) to identify point defects for quantum applications. The models are specifically targeted for use in a prescreening step for accelerated high-throughput workflows, and are therefore designed to avoid the costly relaxation step typically present with ML interatomic potentials (MLIPs). We compare performance for single and double point defect systems in 4H-SiC with ridge, kernel ridge, and multilayer perceptron (MLP) models using three different descriptors representing the defect systems. For vacancies and substitutions, the optimized models give mean absolute errors (MAEs) of 0.437 eV for the formation energy and 0.202 eV for ZPLs, which is just above the level at which such predictions can be useful even beyond the targeted prescreening, i.e., in some applications they may completely replace the need for costly DFT calculations. For interstitials the MAEs are larger, 1.101 eV for the formation energy and 0.230 eV for the ZPL, which, while still useful for prescreening, will not generally be useful for more detailed characterization. Hence, while the results may be further improved by model design and optimization, the models presented in this work are already useful for prescreening in high-throughput characterization of point defects.
ER -