Photolithography, the process of using light to etch features onto surfaces for the manufacturing of computer chips and optical devices, often fails to accurately match designer’s intentions due to tiny inconsistencies in the manufacturing process. Researchers at MIT and the Chinese University of Hong Kong have developed a machine-learning digital simulator in an effort to bridge this gap.
The simulator mimics a specific photolithography manufacturing process using real data gathered from the photolithography system. This results in a more accurate model of how a device would be fabricated. The research team combined this simulator with another to emulate the performance of the fabricated device in different tasks. Together these simulators allow users to create an optical device that better matches its intended design and ensures optimal task performance.
Cheng Zhang, one of the paper’s co-authors, acknowledged the challenges in developing the simulator, particularly the high cost of real data and lack of prior examples, but concluded it was “worth doing”. Zhang argued that real data is “more efficient and precise” than the assumptions made from purely analytical data.
The newly-developed process, called ‘neural lithography’, applies a neural network trained on real data to the physics-based equations which form the basis of photolithography. The network can then adjust to accommodate any specific deviations in the system.
Data for the neural network to learn from is gathered by creating designs that cover a wide range of shapes and sizes using the lithography system. The finished structures are compared with the original designs and this data pairing is used to train the network.
The end system is a combined simulator which enables users to create optimised designs for specific outcomes. The effectiveness of the machine-learning simulator was demonstrated through the process of creating a holographic element. The finished product created using the new machine-learning process closely resembled the original design specification.
The research team aims to develop their algorithms further to model more complex devices and test the system using consumer cameras. In addition, they are investigating how their approach might be adapted for different types of photolithography systems.
The research has been partially funded by the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund and used MIT.nano’s facilities. Research outcomes will be presented at the forthcoming SIGGRAPH Asia Conference.
