Prediction of oxide coating thickness on moving polymer films using quantum cascade laser spectroscopy and machine learning

Nanometer-accurate inline monitoring of oxide coating thickness on polymer films is essential for quality control in roll-to-roll production. However, measuring coatings in the nanometer range on moving polymer films remains challenging because the measurement samples spatially varying polymer thickness and reflective background conditions along the measurement path, in addition to the low optical contrast between oxide and polymer in the mid-infrared. We present a transflection measurement approach that combines quantum cascade laser (QCL) spectroscopy with data-driven spectral analysis. We apply it to moving oriented polypropylene (OPP) and polyethylene terephthalate (PET) films coated with silicon oxide under laboratory conditions to assess its feasibility for future inline monitoring. In transflection mode, reflected intensity is captured across a spectral range covering the Si–O absorption band, enabling measurements on both transparent and opaque films. Two spectral analysis approaches are proposed and compared. The first combines spectral averaging with linear regression, achieving a mean root mean square error (RMSE) of 7.6 nm for OPP and 15.9 nm for PET. The second combines linear regression and Gaussian process regression (LR-GPR) on blockwise-aggregated spectra. For thickness levels represented in the training data, this approach reduces the RMSE to approximately 4nm for both OPP and PET, corresponding to improvements of approximately 50% for OPP and 75% for PET. While the former approach offers simplicity and interpretability, the latter captures nonlinear spectral relationships. However, generalization of LR-GPR to unseen thickness levels remains the central limitation, with RMSE values of approximately 10 nm to 11 nm. The results show potential for oxide coating thickness estimation, while improved generalization and further validation are needed to establish broader applicability.