Photolithography, the technique of etching features onto a surface using light manipulation, is commonly used in the manufacturing of computer chips and optical devices. However, small deviations during the manufacturing process often impact the performance of the finished product. To address this, researchers from MIT and the Chinese University of Hong Kong have leveraged machine learning to create a digital simulator that mimics a specific photolithography manufacturing process.
The digital simulator incorporates data gathered from the photolithography system itself, which enables a more accurate modelling of the system’s manufacturing outcome. Alongside another simulator that emulates the performance of the fabricated device, the technology aims to bridge the gap between design and manufacturing and ensure that the end product aligns more precisely with the initial blueprint.
This technique represents significant potential for a range of applications such as mobile cameras, augmented reality, medical imaging, and telecommunications. The learning pipeline of the digital simulator is based on real-world data, meaning it can be applied to a vast array of photolithographic systems.
However, building this digital simulator isn’t straightforward. The implementation of real-world data isn’t easy due to costs and challenges in coordinating software and hardware. Despite these obstacles, the researchers established and tested various tools and strategies to overcome them. The results were impressive, demonstrating that the use of real data outperformed that of simulated data.
The process of neural lithography, as the researchers have named this technique, operates by establishing a base using physics-based equations. It then incorporates a neural network trained on actual, experimental data from a user’s photolithography system. The neural network learns to compensate for many of the system’s deviations. It’s critical for the data to originate from real-world operations, as artificial data can’t cover all deviations found in real-world circumstances.
The digital lithography simulator consists of two distinct parts: an optics model that illustrates how light is projected onto the device surface and a resist model illustrating the photochemical reaction that etches features onto the surface. The simulator is integrated with a physics-based simulator that predicts the device’s performance on specific tasks, reflecting the whole design-to-manufacturing process.
The researchers tested the technique by fabricating a holographic element depicting a butterfly image when lit. Compared to other techniques, their holographic element produced a near-perfect butterfly closely matching the design. Future research aims to refine algorithms to model more complex devices and test the system using consumer cameras.
This research has received support from the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund. It was conducted using facilities at MIT.nano. The researchers’ findings will be presented at the forthcoming SIGGRAPH Asia Conference.