Machine Learning‐Accelerated Reconstruction of Periodic Nanostructures with X‐ray Fluorescence Spectroscopy Methods

Abstract

Abstract With advancements in the semiconductor industry, the complexity of three‐dimensional (3D) nanostructures becomes higher with continuously decreasing feature sizes. In order to monitor the processing steps, it is crucial to accurately determine the critical dimensions and composition of these nanostructures. Early detection of production malfunctions is crucial for reasonable production yields, as even minor imperfections can heavily impede device performance. Grazing X‐ray fluorescence spectroscopy methods are non‐destructive and element‐specific methods with high sensitivity for nanostructured surfaces. They enable the quantification and localization of elements within the sample, with information depths sufficient also for buried features. Such measurements can be utilized to reconstruct the shape and size of nanostructure features, with simulated data based on finite element method (FEM) Maxwell calculations. However, the computation time of FEM calculations poses a challenge (especially in the X‐ray energy regime), prolonging the reconstruction process and making it impractical for the characterization of multiple samples. To address this issue, a neural network‐based approach is adopted to replace the time‐consuming FEM‐based forward calculations. In this study, the feasibility of utilizing neural networks for nanostructure reconstruction is demonstrated on various nanostructures. The discrimination limit for different model parameters is assessed and compared against conventional FEM calculation results.

Keywords

• finite element method
• machine learning
• nanostructure characterization
• neural network
• X‐ray fluorescence