Researchers at Chalmers University of Technology in Sweden have developed a machine-learning approach that incorporates knowledge of physics to reduce the time needed to design optical components used in fields including nanophotonics, quantum computing, eyeglass lenses and camera lenses.
The research group, led by Philippe Tassin, professor at the Department of Physics and Astronomy at Chalmers, works on optical components at the nanoscale, where light can be controlled in ways that differ from behavior at larger scales. The team uses computer simulations to investigate artificial optical materials that could overcome limitations found in natural optical materials.
The materials under study could be used to make camera and eyeglass lenses thinner, lighter and more efficient. The work may also be relevant to quantum computing. In collaboration with researchers at Chalmers’ Department of Microtechnology and Nanoscience, where Sweden’s first larger quantum computer is being built, the group is examining whether nanostructured materials can control how light travels. One possible application is the transmission of information between quantum computers, or over longer distances, using optical frequencies through mechanically compliant photonic crystals.
The research relies on simulations run on supercomputers, using machine learning and neural networks to predict material properties and guide design. According to the researchers, generating data to train such networks has been a major bottleneck. A single data point can take between 10 minutes and an hour to produce, and as many as 40,000 simulations may be required.
The team said its new method reduces the time needed to generate training data from about 30 days to about three days. The improvement comes from embedding knowledge of electromagnetism and other physical laws into the neural network before training, rather than requiring the system to infer those constraints entirely from simulation data.
“When we fed the super-brain information about the laws of physics, it immediately got much smarter. Our calculations now take one tenth of the time previously required,” Tassin said.
The researchers initially introduced recognizable equations into the network to make its predictions easier to interpret. Testing showed that the approach also reduced the amount of data needed for effective training. Their method is described in an article published in Laser & Photonics Reviews.
“Once we’d trained the network, we could ask it to examine any structure at all and get the optical properties in a millisecond. With these new networks, we get better estimates and avoid obvious errors,” said Viktor Lilja, a doctoral student at Chalmers’ Department of Physics and Astronomy. Philippe Tassin thinks that the time saved is the biggest benefit. “Now that we can work so much faster, we can speed up design development for optical components.”
Image/Graphic/Illustration: Chalmers University of Technology | Anna-Lena Lundqvist
