Feature-driven three-stage hybrid model for thin film parameter estimation from XRR curves

Neural network-based models offer a promising approach for X-ray reflectivity (XRR) prediction, enabling millisecond-scale estimation of thin film parameters. However, their accuracy is often limited by the non-uniqueness of solutions. To address this challenge, we propose a three-stage hybrid network model comprising an XRR curve feature extraction module, a neural network prediction module, and an interior-point method optimization module. Experimental results on Mo/Si bilayer thin films show that, compared to a traditional single-stage network model trained on raw XRR curves, the proposed three-stage model significantly reduces ε R (defined as the mean squared error between predicted and measured XRR curves). Specifically, when evaluated on experimentally measured data, the model lowers ε R to approximately 37.8% at the second stage and further reduces it to around 12.9% at the third stage.