Researchers from MIT and the Chinese University of Hong Kong have leveraged machine learning to construct a digital simulator to enhance the precision of photolithography and bridge the gap between design and manufacturing. Photolithography, a crucial manufacturing process in computer chip production and optical device fabrication, suffers from slight deviations that can lead to shortcomings in the resulting products. The novel simulator developed by the research team factors in authentic data from the photolithography system, enabling it to model the fabrication process with higher accuracy.
Integrated into a design framework, the simulator works in conjunction with another digital simulator that models the final device’s performance in tasks such as image generation with computational cameras. Combined, these simulators facilitate the production of an optical device more true to its design and capable of peak task performance.
The application of this machine learning technique could potentially refine and speed up the production of optical devices in numerous fields, including mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications. The novel digital simulator uses data from real environments, increasing its applicability for a range of photolithography systems.
Titled “neural lithography,” the method developed by the MIT researchers blends physics-based equations and a neural network trained on real data from an actual photolithography system. The approach compensates for specific deviations in the system, enhancing its accuracy. The process of data gathering entails producing a variety of designs with a wide range of sizes and shapes using the photolithography system, measuring the final structures, and comparing results with the original design specifications. This data is utilized in training the neural network for the digital simulator.
The digital lithography simulator has two components: the optics model showing how light is projected on the device surface, and the resist model demonstrating the photochemical reaction to create the surface features. In a downstream task, the learned photolithography simulator is connected to a physics-based simulator predicting the performance of the fabricated component on this task, such as how a lens diffracts light.
The new method was tested through the fabrication of a holographic element that generates a butterfly image when lit. The team’s holographic element produced a butterfly image nearly identical to the design, outperforming other techniques. The researchers also created a multilevel diffraction lens that delivered better image quality than other lenses. The team aims to further refine the algorithms to model more complex devices and test consumer cameras. They also plan to expand their technique to different photolithography systems. The research is supported by several organizations, including the U.S. National Institutes of Health.
