Photolithography is a crucial process in the manufacturing of computer chips and other optical devices, but validity between the design and the final product often falls short due to tiny variations in the manufacturing process. To address this issue, researchers from MIT and the Chinese University of Hong Kong have developed a machine-learning aided digital simulator to mimic the photolithography process with greater precision. This new approach uses real data gathered from the photolithography process to more accurately represent how a design would translate into the final product.
The scientists also incorporated another digital simulator into their design platform that replicates the performance of the manufactured device in various downstream tasks. This means that they can incorporate the inputs from the two simulators to create an optical device that more closely resembles its intended design and performs at an optimal level.
This research has considerable implications in a range of fields, with the potential to improve the accuracy and efficiency of optical devices used in mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications.
Traditionally, real data has been expensive and difficult to coordinate in a way that facilitates the construction of a hi-fidelity dataset. As such, many initial simulations have used approximations and estimations derived from physics equations. These equations provide a baseline understanding of the fabrication process but cannot account for all the specific variations within a particular photolithography system. Consequently, manufactured devices often underperform in comparison to their original designs.
The researchers at MIT refined this approach by integrating physics based equations with a neural network model trained on real, experimental data gathered from a photolithography system. This neural network (a machine-learning model based loosely on the human brain) learns to address many of the system-specific variations that contribute to discrepancies in design implementation.
The researchers generated many designs that could be fabricated using the photolithography system and measured these against their intended specifications. They then trained the neural network using this practical data. The specifics of how a particular optical system performs depends on the data used to train it. Therefore, by using real-world data, it is possible to more accurately predict the nuances and deviations in how a design is realized.
The researchers tested their dual-simulator method by creating a holographic element that produced a near-perfect butterfly image when illuminated. Their multilevel diffraction lens also demonstrated superior image quality when compared to other similar devices.
Future research plans involve enhancing the capabilities of their algorithms to simulate more complex devices, piloting the system using consumer-grade cameras, and expanding compatibility with other types of photolithography systems, such as those using deep or extreme ultraviolet light.
The research was made possible with support from the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund. Much of the research took place using MIT.nano’s facilities.
