Researchers from the Massachusetts Institute of Technology (MIT) and the Chinese University of Hong Kong have developed a digital simulator using machine learning to optimize the manufacturing process used in fabricating devices like computer chips and lenses. This technology, known as photolithography, manipulates light to precisely etch features onto a surface, but minute deviations can result in products that fail to meet design expectations.
The researchers’ simulator, dubbed “neural lithography,” uses real data from the photolithography system to model the manufacturing process more accurately. It is integrated into a design framework alongside another simulator that predicts how the manufactured device will perform in downstream tasks, such as producing images with computational cameras. This approach allows for the creation of an optical device that closely matches its design and performs optimally.
Neural lithography may be used to improve the accuracy and efficiency of optical devices in various applications, including mobile cameras, augmented reality, telecommunications, and medical imaging. It could also be adapted for use with a wide range of photolithography systems, thanks to its data-driven approach.
Cheng Zheng, an MIT graduate student who co-authored the research paper, noted that real-world data can prove more effective than data generated by simulators based on analytical equations. However, such data can be expensive and tricky to gather.
The researchers’ method began with creating a multitude of designs with varying feature sizes and shapes. After fabrication using the photolithography system, the final structures were compared to the original designs. The results were paired and used to train a neural network that factors in many of the system’s specific deviations.
Comprising optics and resist models, the digital lithography simulator captures how light is projected onto the device and how the photochemical reaction etches features onto the surface. The researchers link this model to a physics-based predictor of the device’s downstream performance. The dual simulators are then integrated into a broader framework that advises users on how to create a device design that fulfills specified performance goals.
In testing, the neural lithography technique successfully fabricated a holographic element that produced a nearly perfect butterfly image when illuminated. This outcome more closely aligned with the design than those of devices created using other methods. The system was also able to produce a multilevel diffraction lens with superior image quality.
The researchers plan to enhance their algorithms to model more sophisticated devices, test consumer cameras, and adapt their approach for other types of photolithography systems. The research is funded by the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund, and took advantage of MIT.nano’s facilities. The findings will be presented at the SIGGRAPH Asia Conference.
