Photolithography is a crucial technique in the production of computer chips and optical devices, but it is susceptible to micro discrepancies which can result in the final devices not performing as designed. MIT and the Chinese University of Hong Kong researchers have helped resolve this issue, using machine learning to create a digital simulator that replicates a photolithography manufacturing protocol. The benefit of this simulator is that it uses actual data gathered during the photolithography process, allowing for more precise modeling of how a device would be created.
The researchers incorporated this digital simulator into a design framework along with a second simulator, which mimics the functioning of a constructed device in processes such as image generation from computational cameras. The two interlinked simulators enable users to create an optical device that more accurately aligns to its design and performs optimally.
This innovative technique promises to support the development of more precise optical devices used in mobile cameras, augmented reality, medical imaging, entertainment, and telecommunication applications. A significant advantage of this method is that it can be applied to numerous photolithography systems as its training pipeline uses real-world data.
The use of real data to create the digital simulator is a novel approach, attributed to the cost of data and the lack of established processes for coordinating software and hardware to generate a high-fidelity dataset. However, the research team was confident in trying new characterization tools and data-exploration strategies to come up with an effective plan. The results of this experiment confirmed the advantage of real data over generated data from simulators using analytical equations.
The technique they developed, termed “neural lithography,” employs physics-based equations foundational to creating a photolithography simulator, which is then overlaid with a neural network trained on actual data from a user’s photolithography system. The neural network, modeled loosely around the human brain’s structure, learns to compensate for many system-specific deviations. This method requires many different designs to be exploited, covering a wide range of detailed sizes and shapes. The researchers then compared the final structures against their correspondingly designed specifications and used this data to train the digital simulator’s neural network.
The digital lithography simulator contains two elements: an optics model delineating how light falls onto the device’s surface, and a resist model illustrating how features are formed on the surface during the photochemical reaction. The simulated photolithography protocol is also linked to a physics-based simulator that predicts the device’s functioning. The user determines the desired outcomes for a device, and the simulators unite to guide the user on designing a device that will meet these performance objectives.
The researchers tested this approach by creating a holographic item that generates a butterfly image when light is cast onto it. The holographic item showed a near-perfect resemblance to the original design and demonstrated improved image quality than other devices.
Looking ahead, the team aims to refine their algorithms to emulate more complex devices and test the system using consumer cameras. The researchers also plan to adapt their technique so it can be utilized across different types of photolithography systems, such as those using deep or extreme ultraviolet light. The U.S National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund have supported this research.
