Semiconductors are essential for modern technology as electronic devices highly depend on transistors and integrated circuits . In addition, semiconductors with various properties are required for developing application-specific devices. For instance, SiC is suitable for high-temperature and high-power devices due to its good thermal conductivity and wide bandgap . Various researchers performed high throughput material screening with density functional theory (DFT) to discover potential semiconductors ,,. However, computationally discovering stable semiconductor materials has become challenging because of the difficulty of finding the ground state crystal structures.
As a solution, the researchers from the Machine Learning and Evolution Laboratory at the University of South Carolina, the University of Colombo Materials Informatics research group, and the University of Moratuwa propose a computational pipeline that combines artificial intelligence with DFT-based high throughput calculations. Here, we use CubicGAN, a Generative adversarial network (GAN). GAN learns patterns and distributions from existing data to generate new data. A GAN contains two main components, which are the generator and discriminator. The discriminator learns to identify the real data from the fake data generated by the generator. CubicGAN is the first work to generate materials at a large scale using a GAN . We developed a binary classifier to filter the semiconductors from the generated materials. Finally, the DFT calculations were performed to verify the stability and semiconductor properties of the filtered materials.
The CubicGAN model generates ternary and quaternary cubic materials with 221, 225, and 216 space group symmetries. Therefore, we first analyzed the ternary and quaternary nonmetals (semiconductors/insulators) and metals in the Materials Project (MP) database . As a result, it could be found that around 73% of the 39,024 quaternary compounds are nonmetals, whereas only about 44% of the 63,784 ternary materials are nonmetals. This indicates a high probability of finding semiconductors from the quaternary dataset. As a result, the elemental properties of quaternary materials were further studied to identify the important patterns in the data. It was shown that the probability of finding a nonmetal is high when the percentage of metal (PM) and transition-metal (PTM) elements in the chemical formula is low. The metals:nonmetals ratio is 1:3 at the PM < 60%. It becomes 5:1 when PM >60%. At PTM < 5%, the metals:nonmetals ratio is 1:6. On the contrary, the number of metals is considerably higher than the number of nonmetals when PTM > 30%.
To screen the semiconductors from the materials generated by CubicGAN, we developed a random forest classifier (RFC) and a deep neural network (DNN). Since the CubicGAN generates only the cubic materials, we collected pretty formulas, and bandgap details of quaternary cubic materials from the MP database. Out of 4,016 cubic materials in the dataset, 2,578 materials were nonmetals, and 1,438 were metals. The feature set of the above two classifiers was created considering 55 elemental and electronic structure attributes, such as the first ionization energy and the total number of valence electrons. The weighted average and the maximum difference of those properties associated with a given chemical formula were computed as features. The DNN classifier was developed using two hidden layers, where the first layer includes 200 neurons and the second layer contains 100 neurons. The RFC model was developed with 500 decision trees.
Both models were trained with 98% of the cubic quaternary material dataset, while 2% was used for testing the models. To assess the predictive performance of the models and evaluate how the models perform outside the training samples, we used 10-fold cross-validation. The mean accuracy of the cross-validation for the DNN model is 0.92±0.034, whereas that for the RFC model is 0.88±0.013. However, the RFC model performs slightly better than the DNN model on the testing set. The accuracy of the DNN (RFC) classifier for the test set is 0.88 (0.91).
To show the procedure of finding stable semiconductors, we applied the RFC model to the 323 mechanically and dynamically stable quaternary cubic materials generated by the CubicGAN model. Based on the results, 137 compounds were classified as nonmetals. Using DFT calculations, we discovered 12 compounds with chemical formulas in AA'MH6 form. Those structures have the space group symmetry F-43m (216). They exhibit zero energy-above-hull, indicating thermodynamical stability. Those materials are NaRhH6, BaSrZnH6, BaCsAlH6, SrTlIrH6, KNaNiH6, NaYRuH6, CsKSiH6, CaScMnH6, YZnMnH6, NaZrMnH6, AgZrMnH6, and ScZnMnH6. Kadir et al. reported five different AA'MH6 semiconductors with M = Ir . They experimentally synthesized NaCaIrH6, NaSrIrH6, NaBaIrH6, KSrIrH6, and KBaIrH6 by direct combination of the alkali (Na and K), alkaline earth (Ca, Ba, and Sr) binary hydrides/deuterides with Ir powder. This indicates a high possibility of synthesizing the stable materials discovered by our research.
Our accurate hybrid functional calculations confirm that those 12 materials have bandgaps greater than 2 eV, suggesting those are wide-bandgap semiconductors. The wide-bandgap semiconductors are useful for designing optical devices emitting green, red, and UV frequencies and high-temperature power applications [8,9]. The bandgap analysis shows that BaSrZnH6 and KNaNiH6 are direct-bandgap semiconductors, whereas others have indirect bandgaps. Direct bandgap semiconductors are desired for LED and laser applications. Moreover, in addition, wide-bandgap semiconductors with direct bandgap are widely studied for solar cells because of optical transparency . It could also be shown that BaNaRhH6, KNaNiH6, CaCsMnH6, and NaYRuH6 materials have very flat bands near the Fermi level. Compared to other materials, BaSrZnH6 contains narrow bands near the Fermi level. As a result, this can lower the effective mass of the carriers, and this can affect charge transport. Finally, it can be mentioned that the above stable semiconductors can be used for numerous applications based on the band structure properties.
- Yu, L. et al. High-performance wse2 complementary metal oxide semiconductor technology and integrated circuits. Nano Lett. 15, 4928–4934 (2015).
- Casady, J. & Johnson, R. Status of silicon carbide (sic) as a wide-bandgap semiconductor for high-temperature applications: A review. Solid State Electron. 39, 1409–1422 (1996).
- Ortiz, C., Eriksson, O. & Klintenberg, M. Data mining and accelerated electronic structure theory as a tool in the search for new functional materials. Comput.Mater. Sci. 44, 1042–1049 (2008).
- Luo, S., Li, T., Wang, X., Faizan, M. & Zhang, L. High-throughput computational materials screening and discovery of optoelectronic semiconductors. WIRES Rev.Comput. Mol. Sci. 11, e1489 (2021).
- Li, Y. & Yang, K. High-throughput computational design of organic-inorganic hybrid halide semiconductors beyond perovskites for optoelectronics. Energy Environ. Sci. 12, 2233–2243 (2019).
- Zhao, Y. et al. High-throughput discovery of novel cubic crystal materials using deep generative neural networks. Sci. Lett. 8, 2100566 (2021).
- Jain, A. et al. The Materials Project: A materials genome approach to accelerating materials innovation. APL Materials 1, 011002 (2013).
- Takahashi, , Yoshikawa, A. & Sandhu, A. Wide bandgap semiconductors. (Springer-Verlag Berlin Heidelberg. 239 2007).
- Millan, J., Godignon, P., Perpiñà, X., Pérez-Tomás, A. & Rebollo, J. A survey of wide bandgap power semiconductor IEEE Trans. Power Electron. 29, 2155–2163 (2013).
- Woods-Robinson, R. et al. Wide band gap chalcogenide semiconductors. Chem.Rev. 120, 4007–4055 (2020).