Modeling Electrical Properties of Oxide Glass via Improved Wasserstein Generative Adversarial Network for Data Augmentation

With the advancement of the Materials Genome Initiative and the establishment and improvement of the glass database, data-driven machine-learning methods penetrate the field of glass materials, and remarkable progress is made in the modeling of the physical, mechanical, and optical properties of glass materials. In recent years, oxide glasses have found increasingly widespread applications in the field of electronic applications, and their electrical properties as critical indicators have attracted much attention. However, modeling on the electrical properties of oxide glasses remains a challenge, and insufficient data suppress the model's performance. To address the issue of limited data availability on the electrical properties of oxide glasses, this work was to propose a data augmentation framework named WGAN-GP-CP tailored for glass materials.

Based on the Wasserstein generative adversarial network with gradient penalty (WGAN-GP) model, an additional composition penalty term (CP) was added to the generator's loss function, and a method for evaluating sample quality was proposed, which could assess the quality of the generated samples from three aspects, i.e., diversity, accuracy, and uniqueness. Moreover, the XGBoost algorithm was employed for model training, and the model was interpreted through SHAP single-feature analysis and feature interaction analysis.

The trained generators can generate high-quality synthetic samples, expanding the existing datasets effectively. The results of comparative studies demonstrate that the data-augmented models exhibit mitigated overfitting and significantly enhanced generalization capability. The generalization performances of the room-temperature bulk resistivity, relative dielectric constant, and dielectric loss models increase from 0.784 to 0.838, 0.862 to 0.897, and 0.801 to 0.861, respectively. The results obtained from SHAP analysis are generally consistent with the classical physicochemical understanding of the electrical properties of oxide glasses. In addition, some interesting mixing effects and interactive relationships are also revealed.

The data augmentation framework proposed in this work offered a strategy for developing high-performance data-driven models under data scarcity conditions. The insights from SHAP single-feature analysis and feature interaction analysis could provide an important guidance for a deeper understanding of the relationship between the composition and electrical properties of oxide glasses. This research shifted the optimization of oxide glass electrical properties from empirically driven trial-and-error methods to data-driven scientific design frameworks, laying a foundation for the development of high-performance glass materials in the field of electronic applications.