Data-driven prediction of optical material properties using machine learning

Accurate prediction of optical material properties is essential for the efficient design of photonic, optoelectronic, and energy-related devices. Conventional experimental characterization and first-principles computational methods, while physically rigorous, are often time-consuming and computationally expensive, limiting their applicability for large-scale material screening. In this work, a systematic data-driven machine learning framework is proposed for predicting key optical properties of inorganic materials, including refractive index, optical band gap, and dielectric constant. A dataset comprising 1,894 chemically diverse inorganic compounds was constructed from publicly available experimental and high-fidelity computational databases. Composition-based elemental descriptors were employed to represent material chemistry without requiring explicit crystallographic information. Several supervised regression models were evaluated, with ensemble-based approaches demonstrating superior performance. The best-performing model achieved coefficients of determination (R²) of 0.87, 0.85, and 0.89 for refractive index, optical band gap, and dielectric constant, respectively, with corresponding mean absolute errors (MAE) below 0.15 for all target properties.

Hali tarjima qilinmagan