Silicon carbide is widely used in power electronics thanks to its wide bandgap and high thermal conductivity. Point defects in it can degrade device performance, but at the same time they open a path to quantum computing and communication. The search for suitable defects traditionally relies on high-performance first-principles calculations, which require large computational resources. Machine learning can speed up this process if it learns to predict key defect properties without full structural relaxation.
The authors used a dataset of 39,333 double and 853 single point defects in 4H-SiC taken from the ADAQ database. Each defect was described in three ways: an atomic descriptor (atomic numbers, distances, charge, spin), a descriptor with Wyckoff letters (adds local symmetry), and a descriptor with Wyckoff coordinates (details interstitial positions). Three models were trained: ridge regression, kernel ridge regression with a Gaussian kernel, and a multilayer perceptron. Performance was evaluated by mean absolute error via 10-fold cross-validation.
For vacancies and substitutions, the best mean absolute error was 0.437 eV for the formation energy and 0.202 eV for the zero-phonon line. For interstitials, the errors are higher: 1.101 eV and 0.230 eV, respectively. The atomic descriptor proved best in three of the four tasks, losing out only on the formation energy of interstitials, where Wyckoff coordinates helped. The multilayer perceptron showed the lowest errors for formation energies, and kernel ridge regression for zero-phonon lines. Learning curves indicate that adding data could further improve the results.
The work demonstrates that even simple descriptors without full atomic structure allow predicting defect properties with accuracy sufficient for preliminary screening. This lowers the entry barrier for using machine learning in materials science and opens the way to mass screening of unpromising candidates before launching expensive calculations. The difference from neighboring works is that here no structural relaxation is required, and two quantities are predicted at once: the formation energy and the zero-phonon line.
The models can be improved through more informative descriptors, especially for zero-phonon lines, where errors are still higher. It may be possible to include data on relaxed structures in the training or to use graph neural networks. Further reduction of errors will allow fully replacing first-principles calculations for some classes of defects.
The method will speed up the search for point defects for quantum technologies and power electronics. It is also applicable to other semiconductors with a similar crystal structure.
The next step could be expanding the dataset and testing the models on other silicon carbide polytypes. It is also worth exploring new descriptors that account for the electronic structure of the defect.
The work addresses the fundamental problem of predicting material properties without a full calculation of the electronic structure. It shows that key defect characteristics can be extracted from limited information, which is important for scaling up the search for new materials.
In 4H-SiC there are more than 200 polytypes, but it is 4H that is used in most commercial devices. Point defects in it can emit single photons, making them candidates for quantum cryptography.