Researchers at Massachusetts Institute of Technology (MIT) and Chinese University of Hong Kong have invented a machine learning-based digital simulator to shrink the gap between design intention and actual manufacturing of computer chips and optical devices. The process of photolithography used in creating such devices often leads to tiny deviations between theoretical design and practical production. The new simulator, which is based on real data from the photolithography system, significantly reduces these deviations, optimising the production process.
According to Cheng Zheng, a mechanical engineering graduate at MIT, while this concept may seem straightforward, the lack of precedent in co-ordinating software and hardware to build a high-quality dataset has discouraged previous attempts. His digital simulator is more efficient and precise than those based on analytical equations, despite the expense and the initial uncertainty.
The simulator integrates into a design framework, in conjunction with a performance emulator that illustrates how the actual fabricated device performs tasks, for example producing images with computational cameras. Thus, the simulator allows scientists and engineers to accomplish their desired results more accurately and with improved efficiency.
The impact of this research extends beyond the manufacturing of computer chips and optical devices, resulting in advancements in numerous areas. By harnessing real-world data in the learning process, the simulator bears applications in a range of photolithography systems; from mobile cameras and augmented reality to medical imaging, entertainment, and telecommunications.
The researchers used the photolithography system to generate numerous device designs with different shapes and dimensions to collect data for their method. The system then used these designs to fabricate corresponding devices. Comparing the original specifications to the final structures, the researchers paired the theoretical with the empirical data and used this information to train their neural network.
The trained digital simulator has a dual structure. The first component involves the projection of light onto a device surface. The second simulates the photochemical reaction which results in distinctive features. When integrated, these models allow for accurate and efficient fabrication of optical devices.
Future applications of this method include more complex devices and consumer camera systems. The research team also hope to expand the utility of this tool to serve additional types of photolithography systems; such as those involving 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.
