Photolithography entails manipulating mild to exactly etch options onto a floor, and is often used to manufacture pc chips and optical gadgets like lenses. However tiny deviations through the manufacturing course of typically trigger these gadgets to fall wanting their designers’ intentions.
To assist shut this design-to-manufacturing hole, researchers from MIT and the Chinese language College of Hong Kong used machine studying to construct a digital simulator that mimics a selected photolithography manufacturing course of. Their approach makes use of actual knowledge gathered from the photolithography system, so it might probably extra precisely mannequin how the system would fabricate a design.
The researchers combine this simulator right into a design framework, together with one other digital simulator that emulates the efficiency of the fabricated machine in downstream duties, similar to producing photographs with computational cameras. These related simulators allow a consumer to provide an optical machine that higher matches its design and reaches the perfect job efficiency.
This method may assist scientists and engineers create extra correct and environment friendly optical gadgets for functions like cell cameras, augmented actuality, medical imaging, leisure, and telecommunications. And since the pipeline of studying the digital simulator makes use of real-world knowledge, it may be utilized to a variety of photolithography programs.
“This concept sounds easy, however the causes individuals haven’t tried this earlier than are that actual knowledge may be costly and there are not any precedents for the best way to successfully coordinate the software program and {hardware} to construct a high-fidelity dataset,” says Cheng Zheng, a mechanical engineering graduate pupil who’s co-lead creator of an open-access paper describing the work. “We’ve got taken dangers and executed intensive exploration, for instance, creating and making an attempt characterization instruments and data-exploration methods, to find out a working scheme. The result’s surprisingly good, exhibiting that actual knowledge work rather more effectively and exactly than knowledge generated by simulators composed of analytical equations. Although it may be costly and one can really feel clueless at the start, it’s price doing.”
Zheng wrote the paper with co-lead creator Guangyuan Zhao, a graduate pupil on the Chinese language College of Hong Kong; and her advisor, Peter T. So, a professor of mechanical engineering and organic engineering at MIT. The analysis might be offered on the SIGGRAPH Asia Convention.
Printing with mild
Photolithography entails projecting a sample of sunshine onto a floor, which causes a chemical response that etches options into the substrate. Nevertheless, the fabricated machine finally ends up with a barely totally different sample due to miniscule deviations within the mild’s diffraction and tiny variations within the chemical response.
As a result of photolithography is complicated and arduous to mannequin, many present design approaches depend on equations derived from physics. These basic equations give some sense of the fabrication course of however can’t seize all deviations particular to a photolithography system. This will trigger gadgets to underperform in the actual world.
For his or her approach, which they name neural lithography, the MIT researchers construct their photolithography simulator utilizing physics-based equations as a base, after which incorporate a neural community skilled on actual, experimental knowledge from a consumer’s photolithography system. This neural community, a sort of machine-learning mannequin loosely primarily based on the human mind, learns to compensate for lots of the system’s particular deviations.
The researchers collect knowledge for his or her technique by producing many designs that cowl a variety of function configurations and dimensions, which they fabricate utilizing the photolithography system. They measure the ultimate constructions and evaluate them with design specs, pairing these knowledge and utilizing them to coach a neural community for his or her digital simulator.
“The efficiency of discovered simulators relies on the information fed in, and knowledge artificially generated from equations can’t cowl real-world deviations, which is why you will need to have real-world knowledge,” Zheng says.
Twin simulators
The digital lithography simulator consists of two separate parts: an optics mannequin that captures how mild is projected on the floor of the machine, and a resist mannequin that reveals how the photochemical response happens to provide options on the floor.
In a downstream job, they join this discovered photolithography simulator to a physics-based simulator that predicts how the fabricated machine will carry out on this job, similar to how a diffractive lens will diffract the sunshine that strikes it.
The consumer specifies the outcomes they need a tool to attain. Then these two simulators work collectively inside a bigger framework that reveals the consumer the best way to make a design that may attain these efficiency objectives.
“With our simulator, the fabricated object can get the absolute best efficiency on a downstream job, just like the computational cameras, a promising expertise to make future cameras miniaturized and extra highly effective. We present that, even when you use post-calibration to try to get a greater consequence, it would nonetheless not be nearly as good as having our photolithography mannequin within the loop,” Zhao provides.
They examined this system by fabricating a holographic component that generates a butterfly picture when mild shines on it. When in comparison with gadgets designed utilizing different strategies, their holographic component produced a near-perfect butterfly that extra intently matched the design. Additionally they produced a multilevel diffraction lens, which had higher picture high quality than different gadgets.
Sooner or later, the researchers need to improve their algorithms to mannequin extra sophisticated gadgets, and likewise check the system utilizing shopper cameras. As well as, they need to develop their method so it may be used with various kinds of photolithography programs, similar to programs that use deep or excessive ultraviolet mild.
This analysis is supported, partially, by the U.S. Nationwide Institutes of Well being, Fujikura Restricted, and the Hong Kong Innovation and Know-how Fund.