Abstract: Optical measurement of a sample that includes a structure-of-interest (SOI) optically coupled to an unknown structure is optically measured by extracting from the resulting spectral signal the spectral variation that is correlated to key parameters associated with the SOI and removing the spectral variation from unknown structure that is irrelevant to the key parameters. An offline process is used to generate a physics-based model from a number of calibration measurements using reconstructed spectral signals after removing the spectral variation from the spectral signals that is irrelevant to the key parameters. A machine learning model may be additionally generated using at least the spectral variations correlated to key parameters associated with the SOI. In an in-line process, a sample is measured by filtering the spectral signals from a sample to remove spectral effects from the unknown structure and using the physics-based model or using the trained machine learning model.