Thin-film thickness measurement using physics-constrained adaptive spectral sampling

Abstract Non-destructive thickness characterization of thin films over a wide dynamic range is crucial for advanced optical and semiconductor manufacturing. Conventional broadband reflectance measurement methods based on iterative nonlinear optimization (such as the Levenberg-Marquardt algorithm) often face severe global convergence issues. Specifically, the smooth spectral envelope of ultrathin films contrasts sharply with the dense interference fringes of thick films, making the inverse problem highly prone to getting stuck in local minima. To overcome these limitations, we propose the adaptive feature-fusion thickness predictor (AFTP), a physics-based, data-driven framework for robust and fast initial thickness estimation. AFTP integrates an adaptive sampling mechanism guided by spectral difference gradient to efficiently capture highly nonlinear optical responses. Crucially, the method resolves the characterization degradation inherent in single-modal approaches by fusing spectral-domain reflectance with frequency-domain features via principal component analysis and multi-layer perceptron. Experimental validation demonstrates that, over a broad range of 10–20 000 nm, AFTP limits the absolute initial estimation error to within ±12 nm. By strictly projecting the initial estimate onto a narrow convergence region centered at the global minimum, AFTP avoids computationally expensive grid searches and significantly accelerates the subsequent linearization fitting process. Combined with its exceptional robustness to noise in complex systems, this framework establishes an efficient and reliable method for real-time online metrology.