Researchers in Sweden have developed a machine-learning approach that embeds the laws of physics directly into neural networks.
A new study from Chalmers University of Technology in Sweden shows that machine learning can become far more efficient when it starts with a built-in understanding of the laws of physics. Researchers found that giving an AI system this foundational knowledge dramatically reduced the time needed to develop advanced optical components used in technologies ranging from quantum computers to camera and eyeglass lenses.
“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,” said Philippe Tassin, a professor in the Department of Physics and Astronomy at Chalmers University of Technology.
Tassin’s team works in nanophotonics, a field focused on controlling light at extremely small scales. When light interacts with structures smaller than its wavelength, it can behave very differently than it does on larger scales. However, natural optical materials have limits that restrict how light can be manipulated. To overcome those constraints, the researchers use computer simulations to design artificial optical materials.
These engineered materials could lead to lighter, thinner, and more effective camera and eyeglass lenses. The research may also support future quantum computing technologies. Working with scientists from Chalmers’ Department of Microtechnology and Nanoscience, where Sweden’s first large-scale quantum computer is under development, the team is exploring nanostructured materials that can precisely control the movement of light.
One potential application involves transmitting information between quantum computers, or across longer distances, using optical frequencies and mechanically compliant photonic crystals. These specially designed crystals can reflect light with extremely high efficiency.
Simulations show how to design the material optimally
The researchers rely entirely on supercomputer simulations, using machine learning and neural networks to analyze how different materials behave. These tools help identify material properties and guide the design process.
“I know electromagnetism’s equations inside out and I teach them, but I still can’t draw all the conclusions that the neural network can. The physics is so complex that I don’t understand the properties of a material just by looking at it – but the computer does,” says Philippe Tassin.
Time-consuming to feed data into neural networks
Training neural networks for these simulations has traditionally required enormous amounts of data. Creating a single data point can take anywhere from ten minutes to an hour, and researchers may need as many as 40,000 simulations.
“It might take us a whole month to generate enough data to train the neural network. Then if you realize that you need to add more things, it can take another month,” said Viktor Lilja, a doctoral student in the Department of Physics and Astronomy at Chalmers University of Technology.
The team has now cut that process to about one tenth of the original time. Tasks that once required 30 days can now be completed in roughly three days because the neural network already understands key physical principles before training begins.
Teaching the neural network the laws of physics
The researchers recognized that optical components must always follow the laws of physics and electromagnetism. Instead of forcing the neural network to discover those rules from training data alone, they incorporated the laws directly into the system.
As a result, the AI no longer has to relearn the same physical relationships from scratch each time. The approach emerged while the researchers were trying to make the network’s predictions easier for humans to interpret by embedding familiar equations into the model. During testing, they found that the network also became significantly more capable and required much less training data. The work was described in the journal 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,” Lilja said.
For Tassin, the greatest advantage is the time saved.
“Now that we can work so much faster, we can speed up design development for optical components.”
Reference: “A General Framework for Knowledge Integration in Machine Learning for Electromagnetic Scattering Using Quasinormal Modes” by Viktor A. Lilja, Albin J. Svärdsby, Timo Gahlmann and Philippe Tassin, 17 March 2026, Laser & Photonics Reviews.
DOI: 10.1002/lpor.202502769
The research was funded by the Chalmers Nano Area of Advance, the Swedish Research Council, and the Knut and Alice Wallenberg Foundation. The training of the neural network was carried out using resources provided by the Swedish National Infrastructure for Computing (NAISS) at Chalmers/C3SE and KTH/PDC, in part with funding from the Swedish Research Council. The work was carried out in part within the META-PIX competence centre at Chalmers.
Never miss a breakthrough: Join the SciTechDaily newsletter.
Follow us on Google and Google News.