Photolithography is a manufacturing process that uses light to precisely etch features onto surfaces, such as producing computer chips and optical devices. However, small imprecisions in the process can sometimes result in devices not being produced to specifications. To close this gap, researchers from MIT and the Chinese University of Hong Kong are employing machine learning to simulate the photolithography process and predict its outcome more accurately.
The research team has built a digital simulator integrating real-world data from a specific photolithography system to give a more accurate prediction of the final product. The simulator was then enveloped into a larger design framework, which also included another simulator to imitate the downstream tasks of the fabricated device such as those in computational cameras. This combination of simulators allowed for greater precision in the final product, more closely matching the intended design.
This approach is set to have significant implications across a range of industries, including mobile cameras, augmented reality, medical imaging, entertainment, and telecommunications. The research intends to create more efficient and accurate optical devices, with the machine learning model incorporating real, experimental data and addressing any discrepancies.
The researchers developed their novel technique, termed ‘neural lithography’, by initially using physics-based equations as the foundation of the simulator. This was followed by the introduction of a neural network, a machine learning model reflecting the human brain, which was then trained on real data captured from a user’s photolithography system.
To generate this data, the team created a multitude of designs, featuring a wide specification of sizes and shapes. These were fabricated using the photolithography system, then measured and compared to the design specifications. These paired data were then used to train a neural network, increasing the accuracy of the digital simulator.
The process uses two distinct simulators: an optics model and a resist model, simulating the projection of light on the device’s surface, and the photochemical reaction respectively. These two simulators then collaborate to achieve the desired outcomes within the larger framework, offering suggestions on creating designs that will fulfill the desired performance goals.
The research team tested this technique successfully by fabricating a holographic element that created a precisely accurate butterfly image when light was shone on it. The technique also successfully produced a multilevel diffraction lens of improved image quality.
In the future, the researchers hope to refine the algorithms to model more complex devices and to test the process further using consumer cameras. They also hope to make their approach compatible with different types of photolithography systems. The research was financially supported by the U.S. National Institutes of Health, Fujikura Limited, and the Hong Kong Innovation and Technology Fund.
