Researchers from MIT and the Chinese University of Hong Kong have developed a machine learning technique to bridge the gap between the design and manufacturing processes in photolithography. Photolithography, a technique commonly used in fabricating computer chips and optical devices like lenses, often falls short of the designers’ expectations due to minute deviations during manufacturing. The researchers’ technique, named neural lithography, involves the creation of a digital simulator using machine learning that mirrors a specific photolithography manufacturing process.
The digital simulator uses real data from the photolithography system, allowing it to model the manufacturing process more accurately. The simulator is integrated into a broader design framework, alongside a second simulator that mimics the performance of the fabricated device in tasks such as producing images with computational cameras.
The collaboration of these simulators enables users to create an optical device that is closer to the original design and offers optimal performance for the chosen tasks. Bringing real-world data into the creation of the digital simulator may result in more accurate and efficient devices, potentially benefiting industries such as mobile cameras, medical imaging, augmented reality, entertainment, and telecommunications.
Cheng Zheng, an MIT graduate student and co-lead author of the study, explained that although useful, real-world data can be expensive and there are no existing precedents to guide the effective coordination of software and hardware in creating high-quality datasets.
The researchers’ technique involves developing a photolithography simulator that uses physics-based equations, paired with a neural network trained on experimental data. The neural network learns to account for specific deviations in the photolithography system.
Furthermore, the photolithography simulator developed by the researchers consists of two separate components – an optics model and a resist model. The optics model details how light is projected onto the device surface, while the resist model portrays the photochemical reactions that produce features on the surface. Following the user’s specification of the desired device outcomes, the two simulators work hand in hand within a broader framework, guiding the user in creating a design that aligns with the proposed performance targets.
Testing the technique, the researchers fabricated a holographic element that generated an almost perfect butterfly image in response to light, closely matching the design. They also produced a multilevel diffraction lens with superior image quality compared to other devices.
The researchers plan to further their study by improving their algorithms to model more complex devices, and testing the system using consumer cameras. Additionally, they aim to diversify their approach for compatibility with different types of photolithography systems.
The research was supported in part by the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund. The work was partly conducted using MIT.nano’s facilities.
