Photolithography, a process used to create computer chips and optical devices, can often have tiny deviations during production, causing the final product to fall short of the initial design. To address this, researchers from MIT and the Chinese University of Hong Kong have used machine learning to develop a digital simulator that more accurately models the photolithography process and, therefore, a more precise end product. This solution uses real data collected from the photolithography system and is applied to a design framework which is subsequently integrated with a simulator that mimics the performance of the fabricated device.
This new methodology will allow scientists and engineers to create optical devices, such as mobile cameras, augmented reality equipment and medical implements, with a higher degree of accuracy and efficacy. The research also found that real-world data was more effective than simulating data to train the machine learning model, despite the high costs and complexities associated with collecting real-world data.
The researchers’ approach, named ‘neural lithography’, used physics-based equations as the foundation for their photolithography simulator. They then incorporated a neural network, a type of machine-learning model, trained on real experimental data from the photolithography system. This model allowed for the prediction and mitigation of system-specific deviations.
The photolithography simulator is comprised of an optics model and a resist model, which depict how light projection and photochemical reactions occur, respectively. The model also incorporates a larger framework that shows the user how to make a design that will meet specific performance goals.
The effectiveness of the model was tested by creating a holographic element, to generate an image of a butterfly when lit. The resulting image was found to closely match the initial design, and had superior image quality when compared to images produced using other design methods.
The future aims of the research include enhancing algorithms to model more complex devices and testing the system in real-world scenarios, such as consumer cameras. The researchers also hope to adapt this approach for use with other types of photolithography systems, including those using extreme or deep ultraviolet light.
The research was conducted in part, at MIT.nano’s facilities, and supported by the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund. This research will be presented at the SIGGRAPH Asia Conference.
