Machine Learning-Enabled Optical Property Prediction of Thin Films Using Spectral Data Extraction from Scientific Literature

The integration of machine learning (ML) into materials science has demonstrated significant potential in accurately and efficiently predicting material properties. This study presents a data-driven approach for extracting refractive index spectra from line plots in scientific literature, transforming previously underutilized graphical data into a structured data set for predictive modeling. Advanced data extraction tools, including ChartOCR and Plot2Spectra, were employed alongside rigorous feature engineering to curate a high-quality data set comprising 1,770 samples. The preprocessing pipeline encompassed spectral data extraction, normalization, and the incorporation of key features, such as elemental composition, synthesis parameters, and wavelength-dependent properties. An XGBoost model was trained and evaluated, achieving a mean absolute percentage error (MAPE) of 0.149, highlighting the effectiveness of this approach. These findings underscore the potential of ML-driven methodologies for enhancing optical property predictions and automating data extraction, thereby facilitating broader applications in materials research.