Enhanced prediction and optimization of thin metal film optical properties using optimized ensemble learning models

Thin metal films are essential for expanding sensors, optoelectronic, and photovoltaic technologies. The intricate relationship between material composition, thickness, and production presents significant challenges in optimizing optical properties. The paper introduces an AI-driven framework for the simultaneous prediction and optimization of metal film optical characteristics, such as transmittance(%T), reflectance(%R), and absorptance(%A) using a dataset of 1320 experimentally measured samples across material films of gold, aluminum, nickel, tin, copper, and molybdenum over 200–2000 nm wavelength range. The input features to the model include wavelength and material type. Ensemble Models such as Random Forest, Gradient Boosting, XGBoost, and Extra Trees were trained and optimized through GridSearchCV with stratified K-fold cross-validation. A multitask learning model was also implemented to explore potential improvements from joint prediction. Among all models, the CatBoost Regressor demonstrated superior performance, achieving R² = 0.99928, MAE = 0.21924, and MSE = 0.28203 on average across all outputs. To enhance interpretability, the feature importance analysis was employed, revealing that Material Type had a slightly more predictive influence than Wavelength. Additionally, material-specific error analysis identified challenging prediction zones tied to spectral extremes. The best performing machine learning model was deployed via a web-based GUI, enabling real-time prediction of thin-film optical properties. Overall, the proposed framework provides a scalable, interpretable, and deployable solution for AI-assisted material design and optical characterization. These findings accelerate thin metal film optimization by offering a reliable data-driven route for quick material property identification and enhancement through machine learning.