Design/methodology/approach This study analyzed 35 lubricant samples comprising molybdenum dialkyl dithiocarbamate (MoDTC), zinc dialkyl dithiophosphate (ZnDDP), isooctyl acid phosphorodithioate amine salt (T308) and thiophosphoric ester amine salt (T310A). FTIR spectroscopy yielded extensive spectral data, in which PCA decreased the dimensionality while preserving more than 90% of data variability. Subsequently, an XGBRegressor-stacked ensemble regression model was used to accurately predict the lubricant components. Findings The PCA-SER Model attained high accuracy in predicting lubricant components (R²: 0.983–0.996) for additives such as MoDTC, ZnDDP, T308 and T310A. PCA reduced dimensionality, preserving over 90% variance, while reducing errors to a mean absolute error of 9.9 × 10–5 and a mean squared error of 1.6 × 10–8. These scores illustrate the ability of the model to precisely predict and classify lubricant components, even in complex and high-dimensional FTIR data sets. Originality/value This study presents a novel PCA and stacked ensemble learning framework for analyzing high-dimensional FTIR data, enhancing lubricant classification and prediction, optimizing formulation processes and ensuring quality control of lubricants. Peer review The peer review history for this article is available at: https://publons.com/publon/10.1108/ILT-12-2024-0472/