Analysis of Optical measurements by Machine Learning and Analytical Models for Sensors

Abstract

This thesis evaluates whether machine learning models can surpass a physics-based analytical solver for film thickness retrieval from spectroscopic ellipsometry data under varying noise, resolution, and dataset size conditions. A large synthetic dataset generated via a Fresnel three-layer model is used to benchmark Linear Regression, Random Forest, and Multilayer Perceptron models against a grid-search analytical baseline. While the analytical solver achieves perfect accuracy on clean data due to model consistency, it degrades significantly under non-Gaussian noise, whereas machine learning models remain stable and achieve lower RMSE in noisy conditions. The study shows that dataset size is the primary constraint on accuracy, while spectral resolution can be reduced with minimal performance loss, and that predictive information is concentrated in a limited subset of wavelengths. Overall, the results define the regimes where machine learning outperforms analytical inversion and provide guidelines for efficient ellipsometer design and calibration.