[23p-P01-1~4] 3.1 Basic optics and frontier of optics

[23p-P01-1] Scatterometry using deep learning analysis for isolated scatterer

〇Tetsuya Hoshino¹, Shintaro Narioka¹, Sadao Aoki¹, Masahide Itoh¹, Masami Kobayashi² (1.APE Univ. of Tsukuba, 2.ME Univ. of Tsukuba)

Keywords:scatterometry, fluorescence, deep learning

When the difference in refractive index between the medium and the sample is more than about 0.1, the accuracy of the light wave scattering measurement is improved by two orders of magnitude in the three-dimensional shape measurement as compared with the lens imaging. With this method, the shape of the periodic structure has been already estimated by deep learning, but in this report, deep learning is applied to the isolated shape with high versatility. Based on the optical measurement of oil droplets, ellipses, semi-ellipses, squares, and triangles are discriminated from the cross-sectional shape for calculating the size/aspect ratio. As training data, the width and aspect ratio were changed, and 6072 scattering angle distributions were calculated by rigorous coupled-wave analysis of an isolated system. The accuracy verified with the same test data as the teacher data is 96%. In this estimation, a semi-elliptical shape was obtained for each of the seven oil droplets.