Chongfan Technology
News
21
2025
-
02
AI empowers a new paradigm of hyper-structured optical design
Author:
With the rapid development of meta-optics, the complexity of multi-parameter light field control continues to increase. Traditional design methods face great challenges in handling the coupling relationships of complex multi-parameter optical responses, and currently, there is no universal multifunctional design paradigm and algorithm framework.
Recently,Professor Duan Huigao and Professor Hu Yueqiang from Hunan University, in collaboration with Professor Qiu Chengwei's team from the National University of Singapore,proposed a neural network-assisted end-to-end design framework for full light field control of meta-optics. This framework can quickly generate structural parameters of meta-optical elements from functional target requirements of full parameter control (amplitude, phase, polarization, wavelength). Moreover, this method outperforms traditional forward design in device performance and can break through the limitations of multifunctional channels. The work demonstrated orthogonal polarization-multi-wavelength-multi-depth multiplexed holographic display and non-orthogonal arbitrary polarization-spectral multi-information processing applications. This framework provides a universal paradigm for customized design of multifunctional meta-optical elements and lays the foundation for the future development of computational meta-optics and optoelectronic hybrid devices.
The related results were published under the title "Neural Network-Assisted End-to-End Design for Full Light Field Control of Meta-Optics" in the journalAdvanced Materials.PhD student Chi Hanbin from Hunan UniversityandProfessor Hu Yueqiangare co-first authors of the paper, while Professor Duan Huigao from Hunan University, and Professor Qiu Chengwei from the National University of Singaporeare co-first authors of the paper, while Professor Duan Huigao from Hunan Universityare co-corresponding authors. This work was supported by the National Natural Science Foundation, the National Key Research and Development Program, and related projects from the Hunan Provincial Department of Science and Technology.Multi-parameter light field control is key to achieving advanced optical functions, while traditional methods rely on cascaded large-volume refractive elements, which restrict the development of miniaturized multifunctional optical devices. Meta-optics, with its unique light-matter interaction mechanism and broad design freedom, provides a powerful tool for miniaturized devices for multi-parameter light field control. However, the interdependence between different wavelengths makes the optical response requirements extremely complex, and traditional spatial multiplexing designs face challenges such as low spatial resolution and severe crosstalk.Currently, algorithm-driven meta-optics design has become a research hotspot. However, existing methods are still limited to a few polarization and wavelength functions and lack global optimization and rapid convergence capabilities. For example,as shown in Figure 1,the traditional design strategies for multi-polarization multi-wavelength meta-optics mainly include atomic-level inverse design ({
Figure 1a
}) and system-level separated design ({Figure 1b}). The former is mainly used to optimize periodic meta-units, only meeting 1D spectral curve requirements, making it difficult to adapt to spatial light field control; the latter adopts a multi-step method to design non-periodic meta-optics. Although it has certain multifunctionality, it is prone to introducing local errors, leading to performance bottlenecks, and is still limited to dual polarization and a few wavelength functions. Therefore, there is an urgent need to develop a system-level end-to-end (E2E) design framework ({Figure 1c}) to achieve full light field control of meta-optics while considering global optimization, rapid convergence, and manufacturing simplification.This study proposes a full light field control meta-optics E2E design framework based on multi-algorithm fusion ({Figure 2a}). This framework constructs a differentiable meta-optics forward solver, combined with a neural network-assisted dispersion full parameter Jones matrix and Fourier diffraction propagation, to quickly predict the geometric parameters of meta-units' optical responses to different wavelengths and polarization states. The neural network generates high-precision continuous phase and amplitude responses ({Figure 2b, 2c
}) by learning FDTD simulation data, ensuring the accuracy and differentiability of the forward solver. Based on this, the framework can use automatic differentiation techniques to calculate gradient information and combine it with gradient descent optimization algorithms to continuously iterate and optimize the geometric layout of meta-optical devices until it converges to the target light field for full light field control ({}). This E2E design framework has the advantages of rapid convergence and global optimization, effectively improving design efficiency, avoiding performance losses caused by local optimization, and achieving system-level optimization of high-performance full light field control meta-optical devices.To verify the advantages of E2E design in enhancing meta-optical performance, it was compared with separated design. For dual polarization multi-wavelength holographic multiplexing ({Figure 3a}), as the number of wavelength channels increases, the RMSE value of E2E design is overall lower than that of separated design ({}). This E2E design framework has the advantages of rapid convergence and global optimization, effectively improving design efficiency, avoiding performance losses caused by local optimization, and achieving system-level optimization of high-performance full light field control meta-optical devices.Figure 3b
}), and the reconstructed holographic image profile is clearly visible ({Figure 3c, 3dAs the number of wavelength channels increases, the RMSE value of the E2E design is overall lower than that of the separated design.Figure 3e}). In contrast, E2E design directly optimizes the geometric layout of meta-optical devices with the goal of image quality, effectively avoiding intermediate matching errors. Experimental measurement results ({Figure 3g, 3h}) show that the PSNR values of the holographic images in the X and Y polarization channels of E2E design increased by 4.26 dB and 2.84 dB, respectively, while the SSIM values increased by 6.46 times and 5.78 times, fully demonstrating the excellent performance of E2E design in polarization-wavelength multiplexed holography.To demonstrate the capability of E2E design in breaking through channel coupling limitations, a dispersion full parameter Jones matrix was formed by introducing rotation angles, fully utilizing all design freedoms of rectangular meta-units to achieve orthogonal tri-polarization-multi-wavelength-multi-depth multiplexing ({）。相比之下，E2E设计以图像质量为目标，直接优化超构光学器件的几何布局，从而有效避免中间匹配误差。实验测量结果（图3g，3h）表明，E2E设计的全息图像在X和Y偏振通道的PSNR值分别提升了4.26 dB和2.84 dB，而SSIM值分别提高了6.46倍和5.78倍，充分证明E2E设计在偏振-波长复用全息中的卓越性能。
为展示E2E设计在突破通道耦合限制方面的能力，通过引入旋转角度形成色散全参量琼斯矩阵，充分利用矩形超构单元的全部设计自由度，实现了正交三偏振-多波长-多深度复用（Figure 4a). Traditional separation design is difficult to achieve such complex functions, while E2E design realizes 27-channel multiplexed holography through automatic optimization of the geometric layout of metamaterial optical devices, simulating high image quality with negligible crosstalk (Figure 4b). In the experimental verification, 9 channels were selected for the demonstration of color 3D holography, and noise distribution was introduced during the optimization process, thereby reducing the sensitivity of the holographic image's PSNR value to shape errors, greatly enhancing the robustness of the device (Figure 4c). The fabricated metamaterial optical device (Figure 4d) can generate corresponding color images at different depth planes under different combinations of incident and outgoing polarization: "Rainbow", "Flower", and "Balloon" (Figure 4e-g). E2E design breaks through the coupling relationship between polarization and wavelength channels, greatly expanding the upper limit of metamaterial optical multi-channel multiplexing.
For the complex vector multi-color light field in practical scenarios (Figure 5a), E2E design directly optimizes the geometric layout of metamaterial optical devices by superimposing arbitrary polarization components and considering their dispersion characteristics, simulating 18-channel holographic multiplexing (6 arbitrary polarization states × 3 wavelengths). In the experimental verification, arbitrary polarization multi-wavelength holographic devices were fabricated, capable of displaying Arabic numerals "1" to "6" at different spatial positions, with each number corresponding to a specific polarization state and wavelength (Figure 5c), verifying the device's multi-information display capability. In terms of multi-information detection and imaging, multi-focus superlenses were designed (Figure 5d), whose spectral splitting efficiency curve (Figure 5e) shows a significant intensity peak at the target wavelength, verifying its spectral detection capability. Meanwhile, the polarization contrast curve (Figure 5f) varies between -1 and 1 with changes in the incident polarization state, indicating good polarization detection capability. Further resolution tests (Figure 5g) show that the device exhibits no significant crosstalk during the imaging process, further proving its application potential in incoherent polarization-spectral multiplexed imaging.
This study developed a neural network-assisted end-to-end (E2E) design architecture, establishing a differentiable relationship between geometric parameters and multi-parameter optical response fields, directly optimizing the geometric layout of metamaterial optics using gradient information, achieving high-performance, multifunctional, and reliably manufactured all-optical field control devices. Experimental verification shows that this architecture significantly improves image quality in dual-polarization multi-wavelength holography and breaks through channel limitations, achieving orthogonal tri-polarization multi-wavelength-depth multiplexed color 3D holography. In addition, it further verifies the feasibility of non-orthogonal arbitrary polarization-spectral information processing capabilities in holographic display and detection imaging. Considering non-predetermined functions, it can achieve high-precision multi-task parallel classifiers, showcasing great potential in intelligent applications.
In the future, customized all-optical field control capabilities can provide new solutions for cutting-edge technologies such as all-optical computing and high-dimensional sensing. Moreover, the differentiable characteristics of this architecture enable seamless integration with deep learning training, promising to promote the collaborative optimization of metamaterial optics and artificial intelligence (AI) in the fields of computational optics and optoelectronic hybrid networks. Overall, the system-level E2E design architecture proposed in this study sets a new standard for multi-parameter complex light field control, advancing the application of high-performance metamaterial optical devices in multifunctional integrated advanced optical systems, and opening new possibilities for the future development of optical technology.
Chongqing Chongfan Technology Co., Ltd. Optical Platform Film Thickness Meter OCT Optical Coherence Tomography LabVIEW Control
LATEST NEWS
Thank you for visiting the official website of Chongfan Technology. If you have cooperation intentions or suggestions, please contact us through the following methods, and we will reply as soon as possible, thank you!
Address: Room 403, Building 6, Phase III of R&D, No. 36 Xiyong Avenue, High tech Zone, Chongqing, China.
Telephone: +86-13658337211
E-mail: Sales@cfkeji.net
Website: www.cfkeji.net
Mobile Version
