Photolithography is an important process in the manufacture of computer chips and optical devices like lenses, using light to carve precise features onto a surface. However, minor deviations during the manufacturing process can lead to these devices underperforming when compared to the original designs. To address this issue, researchers from MIT and the Chinese University of Hong Kong have created a digital simulator using machine learning that closely mimics a specific photolithography manufacturing process.
The simulator uses actual data gathered from the photolithography system, thus promising a more accurate representation of how a design would be fabricated. The simulator is integrated into a design framework along with another simulator that estimates the performance of the fabricated device for downstream tasks (such as creating images with computational cameras). This allows the user to design an optical device that aligns more closely with the intended design and delivers the best task performance.
This development could lead to the creation of more precise and efficient optical devices which could find application in mobile cameras, augmented reality (AR), medical imaging, entertainment, and telecommunications. As the learning pipeline for the digital simulator uses real-world data, it can be used for various photolithography systems.
Collecting real-world data for this purpose can be an expensive process and there are no established methods for coordinating software and hardware to build high-fidelity datasets, according to Cheng Zheng, a lead author of the research paper. However, he contends that the benefits of using real-world data far exceed the costs and that the results are “surprisingly good.” The real data helps to more efficiently and accurately model the manufacturing process than data generated by simulators based on analytical equations.
The technique, dubbed “neural lithography,” uses physics-based equations as the foundation for the digital simulator, supplemented by a neural network trained on real, experimental data from a photolithography system. This allows the network to account for specific deviations in the system. The data for the method is collected by fabricating numerous designs using the photolithography system and comparing the final structures against the design specifications. This data is then used to train the neural network used in the digital simulator.
Looking to the future, the researchers aim to refine their algorithms to model more complex devices and to test the system using consumer cameras. They also hope to expand their approach to work with other types of photolithography systems, like those using deep or extreme ultraviolet light. The research is set to be presented at the SIGGRAPH Asia Conference and received funding in part from the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund. The work utilizes MIT.nano’s facilities.
