Photolithography, a process used to etch features onto surfaces like computer chips and optical lenses, often results in devices that underperform due to tiny variations during manufacturing. To address this, researchers from MIT and the Chinese University of Hong Kong have employed machine learning to create a digital simulator that replicates a specific photolithography manufacturing process. This novel technique uses real data gathered from the photolithography system, enabling a more accurate model of how a design would be fabricated.
The simulator, integrated into a design framework, emulates the performance of the fabricated device in downstream tasks, assisting in the creation of more efficient, accurate optical devices for applications like mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications. This approach could be applied to a variety of photolithography systems as it learns from real-world data.
Despite the complexity and high cost associated with handling real data, Cheng Zheng, a mechanical engineering graduate student, asserts the value in developing such a system. He said that real data work far more efficiently and precisely compared to data generated by simulators based on analytical equations.
The new technique, called “neural lithography”, incorporates a neural network into the photolithography simulator. This model, trained on experimental data from the photolithography system, learns to adjust for the system’s particular deviations. The team created many designs with differing feature sizes and shapes using the photolithography system, measured the final structures, and compared them with design specifications. The collected data were used to train the digital simulator’s neural network.
The digital lithography simulator comprises two distinct elements: an optics model showing how light is projected onto the device’s surface, and a resist model demonstrating how the photochemical reaction produces features on the surface. Connected to a physics-based simulator, it predicts the device’s fabricated performance. The two simulators work in tandem within a larger framework, guiding users to achieve desired design results.
Tests involved creating a holographic element that generates a butterfly image when it is illuminated. Comparisons with other techniques revealed that the holographic element created using the digital lithography simulator closely matched the design. Further enhancements, such as considering more complicated devices or employing consumer cameras, are planned for the future. They also aim to extend their method to other types of photolithography systems that use different light spectra.
This research is supported partly by the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund, with some work carried out at MIT.nano’s facilities.
