Chongfan Technology
News
14
2025
-
03
SIGGRAPH Asia | Neural Lithography: Bridging the Gap Between Computational Optical Design and Manufacturing
Author:
Introduction
In recent years, computational optical design technology has rapidly developed, with a large number of optical components being applied to cutting-edge technological fields, including three-dimensional speckle structured light imaging based on holographic diffractive optical elements, and waveguide devices in augmented/virtual reality (AR/VR). Today, computational optics has further expanded to end-to-end computational imaging design, achieving advanced functions such as depth, spectral, and low-illumination imaging through the simultaneous design of diffractive optical elements or metalenses. Compared with traditional optical design, computational optics can provide greater design freedom, creating more powerful optical devices.
However, existing computational optical design methods often ignore the numerical simulation of the actual manufacturing process, leading to a significant gap between the performance of the final manufactured components and the expected performance. At the same time, the consideration of manufacturability is also limited, so the devices often perform poorly in downstream tasks. In this article, the authors define this for the first time as the design-to-manufacturing gap. Because lithography is an important tool for manufacturing optical devices, how to quantitatively model the lithographic errors of real-world systems and how to organically incorporate the upstream lithographic manufacturing process into the downstream design are key to reducing the design-to-manufacturing gap.
To bridge the design-to-manufacturing gap, researchers at MIT and the Chinese University of Hong Kong proposed a digital lithography simulator based on machine learning and physical models that can more accurately simulate the lithographic manufacturing process. They further integrated this simulator with another tool that simulates the performance of downstream tasks into an end-to-end design framework. Through these two cascaded simulators, researchers can manufacture optical devices that are closer to the design and perform optimally in specific tasks.
This technology is expected to help scientists and engineers develop more accurate and efficient optical components in areas such as mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications. It is worth mentioning that because the digital simulator is trained on data from a real lithography system, it can also be applied to the manufacturing processes of other systems or devices.
This achievement was published inSIGGRAPH Asia，the paper is titledBridging the Design-to-Manufacturing Gap in Computational Optics with 'Real2Sim' Learning for Two-Photon Neural Lithography SimulationThe co-first and corresponding authors are Dr. Cheng Zheng of MIT(currently a postdoctoral fellow at Princeton University's Computational Imaging Lab) and Guangyuan Zhaoa PhD student at the Chinese University of Hong Kong; Professor Peter So of MITprovided guidance for this paper.Technical Challenges1. Insufficient high-fidelity lithography simulation
Although existing computational lithography methods using inverse design algorithms are becoming increasingly mature, how to obtain an accurate forward model for manufacturing simulation has been long neglected. High-fidelity modeling of real-world lithography systems is crucial; otherwise, the three-dimensional structure of the actual manufactured devices often differs significantly from the design scheme.
In applications such as computational imaging and display, if the manufacturing process is not considered in conjunction with the subsequent downstream task performance, even if a "perfect" optical diffraction scheme is obtained in the design phase, it may suffer from background noise or other distortions due to manufacturing defects during actual processing, making it difficult to achieve the expected effect.
Figure 1: Key issues facing computational optics (including computational lithography). Left: How to build a digital twin for a real-world lithography system; Right: How to effectively connect downstream task design with upstream manufacturing processes
Innovation and Highlights
The research team proposed
andReal2Sim2Realrespectively addressing the two major challenges.Real2Sim1. Real2Sim: Redefining the computational lithography framework with real dataSim2RealTraditional computational lithography methods, such as optical proximity correction (OPC), rely heavily on physical equations but have difficulty fully capturing deviations in specific lithography systems. To this end, the researchers combined real-world data collected from specific lithography systems with physical equations to train a neural network, constructing a "Neural Lithography" simulator. It consists of an optical model and a photoresist model. By measuring and comparing the actual lithography process and results, it can more accurately reproduce the real behavior of the lithography system.
Figure 2: Real2Sim—Learning the mapping relationship from input layout to processing results from real lithography system data
2. Sim2Real: End-to-end collaborative optimization using AI simulation
In downstream tasks, the research team proposes, for the first time, to cascade the above-mentioned lithography simulator with another simulator based on physical optics and image post-processing, using automatic differentiation to predict the performance of the device in specific applications (such as diffractive imaging lenses and AR/VR waveguides). Under the end-to-end joint design framework, the two simulators iterate and optimize for the user-specified target performance, taking into account both device performance and manufacturability.
Figure 3: Sim2Real—Cascading upstream lithography simulation and downstream performance simulation, enabling end-to-end optimization of manufacturability and task performance.
Experimental Verification
1. Improving the fidelity of existing lithography system simulations
2. Bridging the design-to-manufacturing gap
图4：不同模型的正向光刻仿真精度对比
2. 弥合设计与制造的鸿沟
After improving the accuracy of lithography simulation, the research team designed and fabricated diffractive holographic elements (HOE) and multilevel diffractive lenses (MDL), and integrated a neural lithography simulator into the design process. Results show that by incorporating manufacturing constraints during the design phase, neural lithography successfully narrowed the gap between design and manufacturing:
-
In HOE design, using the neural lithography simulator can generate holograms with higher image quality and significantly improve indicators such as contrast, SSIM, and PSNR.
-
In MDL design, the neural lithography method improved the contrast of direct imaging applications, making the point spread function (PSF) brighter and more concentrated, and enhancing the high-frequency information recovery capability in computational imaging. It also leads to superior performance of the device in downstream tasks (such as computational photography). Improvements from post-processing traditional calibration still do not match the results obtained by incorporating real lithography simulation into the process at the design stage.
Figure 5: Shows a performance comparison of the designed HOE, where the neural lithography method effectively bridges the manufacturing and design gap.
Figure 6: MDL comparative experiments show that neural lithography has significantly improved performance in both direct imaging and computational imaging.
"Neural lithography" provides a feasible path to bridge the gap between computational optical design and actual manufacturing. By integrating the learned digital twin of the lithographic process into the end-to-end design process, the finally manufactured optical components are closer to the expected performance. With the advancement of research, neural lithography is expected to accelerate the R&D and commercialization of advanced computational optical systems in various applications.
This technology not only has implications for the optical field and computational optics, imaging display, and computational lithography, but also provides a reference for other precision manufacturing. In the future, the research team will continue to explore higher-precision algorithms and measurement methods, and expand this method to more manufacturing systems, providing a more comprehensive solution for the design and production of new generation optical and electronic devices.
Paper Project Homepage:
https://neural-litho.github.io/
Paper URL:
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
