Photolithography involves the manipulation of light to precisely etch features into a surface and is commonly used to make computer chips and optical devices such as lenses. But small deviations during the manufacturing process often cause these devices to fail to meet their designers' intentions.
To help bridge this gap between design and manufacturing, researchers at MIT and the Chinese University of Hong Kong used machine learning to build a digital simulator that mimics a specific photolithography manufacturing process. Their technique uses real data collected from the photolithography system, so it can more accurately model how the system would fabricate a design.
The researchers integrate this simulator into a design framework, along with another digital simulator that emulates the performance of the manufactured device in downstream tasks, such as producing images with computational cameras. These connected simulators allow the user to produce an optical device that best fits their design and achieves the best task performance.
This technique could help scientists and engineers create more precise and efficient optical devices for applications such as mobile cameras, augmented reality, medical imaging, entertainment and telecommunications. And because the digital simulator learning process uses real-world data, it can be applied to a wide range of photolithography systems.
“This idea sounds simple, but the reasons people haven't tried it before is that real data can be expensive and there is no precedent for how to effectively coordinate software and hardware to build a high-fidelity data set.” says Cheng Zheng, a mechanical engineering graduate student who is co-senior author of a open access document describing the work. “We have taken risks and done extensive exploration, for example, developing and testing characterization tools and data exploration strategies, to determine a roadmap. The result is surprisingly good and shows that real data works much more efficiently and accurately than data generated by simulators composed of analytical equations. Although it can be expensive and one may feel disoriented at first, it is worth doing.”
Zheng wrote the paper with co-author Guangyuan Zhao, a graduate student at the Chinese University of Hong Kong; and his advisor, Peter T. So, professor of mechanical engineering and biological engineering at MIT. The research will be presented at the SIGGRAPH Asia Conference.
Light printing
Photolithography involves projecting a pattern of light onto a surface, causing a chemical reaction that etches features into the substrate. However, the manufactured device ends up with a slightly different pattern due to minuscule deviations in light diffraction and small variations in the chemical reaction.
Because photolithography is complex and difficult to model, many existing design approaches are based on equations derived from physics. These general equations give insight into the manufacturing process, but cannot capture all deviations specific to a photolithography system. This can cause devices to underperform in the real world.
For their technique, which they call neural lithography, the MIT researchers build their photolithography simulator using physics-based equations as a basis and then incorporate a neural network trained with real experimental data from a user's photolithography system. This neural network, a type of machine learning model loosely based on the human brain, learns to compensate for many of the system's specific deviations.
The researchers collect data for their method by generating many designs covering a wide range of feature sizes and shapes, which they fabricate using the photolithography system. They measure the final structures and compare them to the design specifications, match that data and use it to train a neural network for their digital simulator.
“The performance of learned simulators depends on the input data, and artificially generated data from equations cannot cover real-world deviations, so it is important to have real-world data,” Zheng says.
Dual simulators
The digital lithography simulator consists of two separate components: an optical model that captures how light is projected onto the surface of the device and a resist model that shows how the photochemical reaction occurs to produce features on the surface.
In a subsequent task, they connect this learned photolithography simulator to a physics-based simulator that predicts how the fabricated device will perform in this task, for example, how a diffractive lens will diffract light incident on it.
The user specifies the results they want a device to achieve. These two simulators then work together within a larger framework that shows the user how to make a design that will achieve those performance goals.
“With our simulator, the manufactured object can obtain the best possible performance in a downstream task, such as computational cameras, a promising technology to make future cameras miniaturized and more powerful. We showed that even if post-calibration is used to try to get a better result, it will not be as good as having our photolithography model in the circuit,” adds Zhao.
They tested this technique by making a holographic element that generates a butterfly image when light falls on it. Compared to devices designed using other techniques, its holographic element produced a near-perfect butterfly that more closely resembled the design. They also produced a multilevel diffraction lens, which had better image quality than other devices.
In the future, the researchers want to improve their algorithms to model more complicated devices and also test the system using consumer cameras. Additionally, they want to expand their approach so that it can be used with different types of photolithography systems, such as systems that use deep or extreme ultraviolet light.
This research is supported, in part, by the US National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and technology Fund.
