Semiconductors are fundamental to modern electronics, finding applications in transistors, LEDs, integrated circuits, and solar cells. While traditional semiconductors like silicon and germanium are widely used, the need for materials with diverse properties (e.g., high thermal conductivity, wide bandgap) drives the search for novel semiconductors. High-throughput screening using first-principles calculations has emerged as a powerful tool for discovering new materials, with successful examples including the identification of Cu-In-based halide perovskites as potential photovoltaic absorbers. However, the vastness of the chemical space makes efficient exploration challenging. Generative adversarial networks (GANs) offer a promising approach to address this challenge by learning the underlying patterns of material properties and generating novel structures. Existing GAN-based approaches for material generation have limitations, often focusing on simple systems or specific material families. CubicGAN, in contrast, offers a large-scale approach to generating cubic materials, forming the foundation of this study.

The literature review highlights previous high-throughput screening efforts to discover optoelectronic semiconductors. Studies by Setyawan et al., Ortiz et al., and Zhao et al. demonstrated the utility of high-throughput screening and data mining for identifying materials with specific bandgap properties for applications like radiation detection and solar energy harvesting. The work of Li et al., based on a large dataset of hypothetical materials, proposed candidates for light-emitting and solar cell applications. Regarding GANs in material science, the paper notes the challenges in applying GANs to crystal structures due to the complex and diverse nature of crystal structures and the need to ensure generated structures adhere to physical and symmetry constraints. The authors mention previous work like CrystalGAN and Kim et al.'s work, emphasizing the limitations of scale and focus on specific material families. CubicGAN is presented as a significant advancement, enabling large-scale material generation.

The methodology involves a three-stage process. First, CubicGAN generates a large number of quaternary cubic materials. These materials are then filtered using a binary classifier trained to distinguish between semiconductors/insulators and metals. The classifier uses a feature set of 119 elemental and electronic structure properties, including ionization energy, atomic volume, and the number of valence electrons in different orbitals. Two machine learning models, a deep neural network (DNN) and a random forest classifier (RFC), are trained and compared, with the RFC also employed for feature importance analysis. Finally, density functional theory (DFT) calculations are performed using the Vienna ab initio simulation package (VASP) to validate the thermodynamic stability and characterize the electronic properties of the selected semiconductor candidates. These DFT calculations encompass structure optimization, band structure calculation, density of states analysis, mechanical property calculations (elastic constants, bulk modulus, shear modulus, Young's modulus, Poisson's ratio), phonon dispersion calculations to confirm dynamic stability, and Bader charge analysis to understand charge transfer between elements. The GGA-PBE functional was used for exchange-correlation interactions.

The study identified 12 thermodynamically stable AA'MH6 semiconductors in the F-43m space group: BaNaRhH6, BaSrZnH6, BaCsAlH6, SrTlIrH6, KNaNiH6, NaYRuH6, CsKSiH6, CaScMnH6, YZnMnH6, NaZrMnH6, AgZrMnH6, and ScZnMnH6. DFT calculations revealed that these materials are wide-bandgap semiconductors. Notably, BaSrZnH6 and KNaNiH6 are direct-bandgap semiconductors, while the others are indirect-bandgap semiconductors. The mechanical properties (elastic constants, bulk modulus, shear modulus, Young's modulus, Poisson's ratio) were calculated using DFT, revealing differences between Mn-based materials and NaYRuH6, which exhibit higher stiffness and compression resistance. The phonon dispersion curves confirm dynamic stability, indicating that these materials are not only thermodynamically but also dynamically stable. Bader charge analysis provided insights into charge transfer between elements within the AA'MH6 compounds. The specific heat capacity (Cp) as a function of temperature was also investigated, showcasing differences in Cp between the Mn-based materials and NaYRuH6. The band structures, calculated using DFT, visualize the electronic band gaps and the nature of the band gap (direct or indirect) for each material.

The findings demonstrate the effectiveness of combining GAN-based material generation with high-throughput DFT calculations and machine learning-based screening for the discovery of novel stable semiconductors. The identification of 12 new stable AA'MH6 semiconductors expands the library of potential semiconductor materials, offering possibilities for applications where specific bandgap properties and mechanical properties are required. The observed differences in the properties of AA'MnH6 and NaYRuH6 compared to other AA'MH6 semiconductors suggest potential for tuning the thermoelectric properties of these materials. The d-orbitals of the transition metal atoms at the M-site play a dominant role in the valence region near the Fermi level, highlighting the importance of transition metal choice in determining electronic properties. This study showcases the potential of this integrated computational approach for accelerated materials discovery across various domains.

This research successfully integrated GANs, machine learning, and DFT calculations to discover 12 new stable AA'MH6 semiconductors. The study demonstrated the effectiveness of CubicGAN for large-scale material generation, the capability of the developed classifier to efficiently screen materials, and the power of DFT for validating thermodynamic and dynamic stability and characterizing electronic and mechanical properties. Future work could focus on exploring the thermoelectric potential of these materials based on their tunable properties and investigating the synthesis pathways for experimental verification.

The study focused on cubic materials within a specific chemical space (AA'MH6). The accuracy of the DFT calculations and the classifier performance rely on the underlying approximations and training data used. Further experimental investigation is needed to confirm the predicted properties and assess the real-world applicability of these materials. The applicability of the findings may be limited to the specific space group and chemical compositions studied. Future work might investigate broader chemical spaces and crystal structures.