Enhanced thin-film measurement via integrated Fast Fourier Transform and machine learning

Abstract

Thickness measurement of thin films based on optical technique is distinguished for its non-destructive and highly precise assessment of samples. Traditional measurement algorithms based on reflectometer calculator, including the extremum methods, etc., predominantly exploit the variation in oscillation frequencies associated with different film thicknesses. However, these conventional approaches can significantly increase computational time and complexity, particularly when conducting high-throughput measurements. Recent advancements in machine learning, especially supervised learning algorithms, have demonstrated their capability to accurately construct input-to-output frameworks, making them suitable for spectral-based film thickness measurements. Drawing inspiration from traditional Fourier transform methods, we utilized the Fast Fourier Transform (FFT) for feature extraction and explored two machine learning algorithms, fine-tuning their parameters to enhance performance. Both algorithms consistently achieved high accuracy in regression predictions of film thickness, with FFT-SVR (Support Vector Regression) achieving an R-squared of 0.9995. This robust, end-to-end supervised machine learning algorithm holds great promise for broad application in spectral-based film thickness measurements across both research and commercial settings.