Photolithography, a process of etching detailed patterns onto surfaces using light, is a crucial technique in the design and production of computer chips and other optical devices, such as lenses. However, minute deviations during manufacturing can cause a discrepancy between the designer’s intentions and the actual produced device.
To help bridge this gap between design and execution, a collaborative effort of researchers from MIT and the Chinese University of Hong Kong has employed machine-learning to build a digital simulator. The digital simulator is designed to mimic a specific photolithography manufacturing process and can provide nearly accurate model of realistic device fabrication.
The simulator was created by gathering real data from the photolithography system. This data was integrated into a design framework along with a secondary digital simulator, which imitates the performance of the fabricated device. The coordinated simulators thereby allow a user to produce a device that more closely matches the initial design and achieves optimal performance. This technique could greatly improve the efficiency and accuracy of optical devices, with applications for mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications.
The process of gathering real data for the simulator may be costly and there may be challenges in coordinating the software and hardware, stipulated Cheng Zheng, a mechanical engineering graduate student who co-authored a paper on the research. However, Zheng confirms that the investment is worthwhile. The primary reason is that data generated from these real-life scenarios function more effectively and accurately than data generated by simulators solely driven by analytical equations.
The technique employed by the research team, labelled as neural lithography, uses physics-inspired equations to lay foundational knowledge for the photolithography simulator. The simulator then learns to accommodate for specific deviations of a user’s photolithography system, through a neural network trained with real data from the said system. This method involves substituting artificial data from equations with real-world data ensuring that real-world deviations are accommodated for.
The digital lithography simulator has two parts: an optics model that captures light projection onto the device surface, and a resist model monitoring the photochemical reaction to etch features onto the surface. The two models can then work together within a larger framework to advise the user on how to design a device that can meet performance objectives.
The MIT researchers tested this technique by creating a holographic element that revealed a butterfly image when exposed to light. According to their results, their device showed a butterfly image that closely resembled the initial design, further confirming their method’s validity.
In the future, researchers aim to enhance their procedures to model more complex devices and put their system to the test in consumer cameras. They also wish to broaden their approach to extend to other forms of photolithography systems that use different types of ultraviolet light.
Primary funding for the research came from the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund.
