Researchers at Chalmers University of Technology have developed a machine learning system that learns the laws of physics before training. This makes it possible to design advanced optical materials up to 10 times faster than traditional methods.
This breakthrough could accelerate the development of optical components used in applications ranging from quantum computing to cameras and eyeglass lenses.
“Once we fed the superbrain with information about the laws of physics, it immediately became much smarter.
Our calculations now take a tenth of the time that was previously required,” said Professor Philippe Tassin from the Department of Physics and Astronomy.
Advanced optical material design
The Chalmers team works in nanophotonics, a field focused on controlling and manipulating light on scales smaller than wavelengths.
In these dimensions, light behaves differently than in traditional optical systems, allowing scientists to create artificial materials with properties not found in nature.
Researchers are using supercomputer simulations to design optical materials that can be used to make camera and eyeglass lenses lighter, thinner and more effective. Their research may also support future developments in quantum technology.
The team is working with researchers at Sweden’s Department of Microtechnology and Nanoscience, where the first large-scale quantum computer is being built, to investigate whether it is possible to design nanostructured materials that control the way light travels.
The concept involves using mechanically compliant photonic crystals to transmit information between quantum computers or over long distances using optical frequencies.
Simulations play a central role in this research, helping researchers determine how the materials should be structured to achieve desired optical properties.
Resolving bottlenecks
The research relies heavily on machine learning and neural networks that analyze vast amounts of simulation data to predict how materials will behave.
“I know and teach electromagnetic equations inside and out, but I still can’t draw all the conclusions that a neural network can draw. Physics is so complex that you can’t understand the properties of a material just by looking at it, but computers do,” Tassin said.
However, generating enough data to train these neural networks has traditionally been a time-consuming and resource-intensive process.
According to the researchers, creating a single data point can take 10 minutes to an hour, but a complete dataset may require up to 40,000 simulations.
To address this problem, researchers embedded the fundamental laws of electromagnetism directly into neural networks.
Rather than forcing the system to discover these principles from scratch, the model begins by understanding the behavior of light and electromagnetic fields.
The idea arose as the team sought to make neural network predictions easier to interpret by incorporating equations that would be familiar to researchers.
During testing, we found that this approach also significantly improved the efficiency of the model.
“Once we have trained the network, we can ask it to inspect any structure and retrieve its optical properties within milliseconds. With these new networks, we get better estimates and avoid obvious errors,” said Viktor Lilja.
This improvement reduces simulation data generation time from 30 days to just three days, allowing researchers to accelerate the development of next-generation optical components.
