Data-driven analysis and visualization of dielectric properties curated from scientific literature

Data-driven methods are powerful tools for understanding and discovering materials. However, in certain domains of materials science, the lack of available datasets significantly restricts the research scope. To address this issue, we utilized Starrydata2 web system to compile a comprehensive dataset on dielectric materials, collecting experimental data on over 20,000 samples across a wide compositional space. This dataset enabled the development of machine learning models with high predictive performance and facilitated the identification of important descriptors through recursive feature eliminations. Since the models worked as complete black boxes and hindered intuitive understanding, we employed additional techniques such as dimensionality reduction and clustering to visualize compositional landscape and trends in dielectric properties. By combining the identified important factors with material clustering, we attempted to visualize the effect of crystal lattice on dielectric permittivity within ABO₃ systems, revealing a roughly linear relationship. Our preliminary analyses and visualizations demonstrate the potential of the Starrydata dielectric dataset collected in this study, offering an important foundation for advanced data-driven materials research. This study presents a comprehensive dielectric database curated from the scientific literature with data analyses and visualizations demonstrating its potential, offering an important foundation for advanced data-driven materials research.