image: a. The top space is the data space; the bottom space is the parameter space. The arrows are flow paths to describe the steps of algorithms. The fluctuations in parameter space stand for the difference between the detected response and theoretical response at a local point. b. Schematic layout of the home-made ARS. Sample profile of the grating in the top inset. BS beam splitter, L1, L2 lenses, M mirror, LS light source. c. Generation process. d. Correlations between the results of forward-mapping algorithm (green marks) and AFM measurements (gray line). view more
Credit: by Tongyu Li, Ang Chen, Lingjie Fan, Minjia Zheng, Jiajun Wang, Guopeng Lu, Maoxiong Zhao, Xinbin Cheng, Wei Li, Xiaohan Liu, Haiwei Yin, Lei Shi and Jian Zi
Inverse scattering problems (ISPs) arise in many fields of science and engineering such as computed tomography, fiber Bragg gratings, and optical metrology. However, it still faces major challenges when the parameter range is growing and involves inevitable experimental noises. Many algorithms and measuring techniques have been developed to solve ISPs with good injectivity and stability. However, the existing algorithms are usually time-consuming due to the global optimization of a huge parameter space. For measuring techniques, it is still a challenge to perform a rapid stable high-throughput measurement by a single-shot imaging technique.
In a new paper published in Light Science & Application, researchers from PBG group of Fudan university and Shanghai Engineering Research Center of Optical Metrology for Nano-fabrication (SERCOM) develop a high-throughput Fourier-optics-based angle-resolved imaging spectroscopy (ARS) embedded with robust NN-based algorithms to solve ISPs. Using the home-made ARS, they experimentally
obtain the dispersion patterns with the all-fixed light path by single shot imaging. The forward-mapping neural network (NN) is trained to photonic dispersion patterns, and a gradient-based optimization is further performed on the parameter space to find the optimal solution. When armed with the NN-based algorithms, the reconstructed geometric parameters achieve a strong linear correlation (R2 > 0.982) with the measurements of atomic force microscopy.
“Our home-made ARS enabled us to measure photonic dispersion that contains both abundant band structures features and reflectance information with the all-fixed light path by single shot imaging for solving ISPs.”
“For algorithm, We proposed two solving steps for ISPs of complex objective: (1) building a quick and differentiable forward mapping; (2) gradient-based optimization algorithm. In our work, we used neural network to build the forward mapping. Besides, inherent differentiability of neural networks makes it easy to obtain the gradient information for optimization.”
“Our strategy has made good predictions against actual noises in accord with the AFM measurements, but with nondestructive nature, which means it could provide a versatile methodology to reconstruct grating profiles as well as other ISPs.” they add.