Thin-film thickness measurement is essential in advanced device manufacturing processes to ensure product performance and quality. Spectroscopic reflectometry is widely used in industrial environments because it enables rapid thin-film thickness measurement with a relatively simple system configuration. However, conventional spectroscopic reflectometry relies on theoretical models that assume uniform material properties and smooth surfaces, conditions that are difficult to satisfy in real manufacturing environments.
In particular, in roll-to-roll processes, unintended random surface patterns can be generated due to roller surface conditions, contact-induced pressure, and bending effects. Since it is practically impossible to accurately model such surface conditions mathematically for every measurement, significant discrepancies often arise between measured spectra and theoretical spectra, leading to reduced reliability in thickness analysis.
To address this issue, this study develops a deep learning–based analysis method trained on theoretical reflectance spectra that incorporate surface roughness effects. Analysis of interference spectra obtained from actual experiments demonstrates that the proposed deep learning approach is capable of interpreting cases that could not be analyzed using conventional model-based methods due to severe spectral mismatch. As a result, under identical measurement conditions, the number of analyzable datasets increased by approximately 5.3 times.
