Researchers from MIT and the Chinese University of Hong Kong have developed a machine learning-powered digital simulator that can accurately replicate a particular photolithography manufacturing process. Photolithography is a technique used to intricately etch features onto surfaces, often used in the creation of computer chips and optical devices. Despite its precision, tiny deviations in the manufacturing process can cause the final product to stray from the original design.
These researchers have attempted to close this gap by creating a simulator that uses real data collected from a photolithography system to more accurately predict how a design will be fabricated. The simulator feeds into a large design framework, alongside another digital simulator that predicts how the produced device would perform in subsequent tasks, for instance, image processing in computational cameras.
The use of real-world data in this process means the technique could be applied to an array of photolithography systems, potentially improving the production of optical devices for use in mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications.
Current photolithography design methodologies tend to rely on equations derived from physics. While this offers an overall sense of the process, it can’t account for specific deviations, leading to devices underperforming. The new method, known as neural lithography, builds upon these physics-based equations and integrates a neural network trained on real experimental data. This allows the model to account for system-specific variations.
This enhanced simulator consists of two components: an optics model predicting how light is projected onto the device surface and a resist model highlighting the photochemical reactions that etch features on the surface. These two simulators align within a larger framework, providing guidance on how to design a product capable of achieving the specified outcomes.
The researchers tested this technique by creating a holographic component that displays a butterfly image when exposed to light, which closely matched the original design. They also produced a multi-level diffraction lens that displayed superior image quality.
Going forward, the team intends to refine their algorithms to model more complex devices and to assess the system’s operation in consumer cameras. They also aim to adapt their approach for different types of photolithography systems, including those employing deep or extreme ultraviolet light.
The research was partly conducted using facilities at MIT.nano, and received funding support from the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund.
