Researchers in Sweden have developed a machine learning approach that embeds the laws of physics directly into neural networks.
new research from Chalmers University of Technology This is shown in Sweden machine learning You can be much more efficient if you start with an understanding of the laws of physics. Researchers have found that equipping AI systems with this foundational knowledge can significantly reduce the time needed to develop advanced optical components used in technologies ranging from quantum computers to cameras and eyeglass lenses.
“Once we fed the superbrain with information about the laws of physics, it immediately became smarter. It took 10 times less time to perform calculations than before,” said Philippe Tassan, 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 very small scales. When light interacts with structures smaller than its wavelength, it can behave very differently than at larger scales. However, natural optical materials have limitations that limit how light can be manipulated. To overcome these limitations, researchers are using computer simulations to design artificial optical materials.
These artificial materials could lead to lighter, thinner, and more effective camera and eyeglass lenses. This research may also support future quantum computing technology. Working with scientists at Chalmers Microtechnology and Nanoscience, where Sweden’s first large-scale quantum computer is being developed, the team is exploring nanostructured materials that can precisely control the movement of light.
One potential application is to use optical frequencies and mechanically compliant photonic crystals to transmit information between quantum computers or over long distances. These specially designed crystals reflect light with extremely high efficiency.
Simulations show how to optimally design materials
Researchers rely entirely on supercomputer simulations that use machine learning and neural networks to analyze how different materials behave. These tools help identify material properties and guide the design process.
“Although I know and teach electromagnetic equations thoroughly, I still cannot draw all the conclusions that a neural network can draw. Physics is so complex that you cannot understand the properties of materials just by looking at them, but computers do,” says Philippe Tassan.
It takes a long time to feed the data to the neural network
Training neural networks for these simulations has traditionally required vast amounts of data. Creating a single data point can take 10 minutes to an hour, and researchers may need as many as 40,000 simulations.
“It might take a full month to generate enough data to train a neural network, and then if you realize you need to add something more, it might take another month,” says Viktor Lilja, a doctoral student in the Department of Physics and Astronomy at Chalmers University of Technology.
The team has now reduced the process to about one-tenth of what it originally was. Because neural networks already understand important physical principles before they begin training, tasks that once took 30 days can now be completed in approximately three days.
Teaching neural networks the laws of physics
Researchers recognized that optical components must always obey the laws of physics and electromagnetism. Rather than having a neural network discover rules from training data alone, we built the laws directly into the system.
As a result, the AI no longer needs to relearn the same physical relationships from scratch every time. This approach emerged as researchers sought to make the network’s predictions easier for humans to interpret by embedding familiar equations in their models. During testing, we found that the network’s power was also significantly improved, requiring much less training data. The study was published in the journal Laser & Photonics Reviews.
“Once we have trained the network, we can ask it to inspect any structure and retrieve its optical properties within milliseconds. These new networks give us better estimates and avoid obvious errors,” Lilja says.
For Tassin, the biggest benefit is time savings.
“We can now speed up the design and development of optical components by working much faster.”
References: “A general framework for knowledge integration in machine learning for electromagnetic scattering using seminormal modes” by Viktor A. Lilja, Albin J. Svärdsby, Timo Gahlmann, and Philippe Tassin, March 17, 2026. Laser and Photonics Review.
DOI: 10.1002/lpor.202502769
This research was funded by the Chalmers Nano Area of Advance, the Swedish Research Council, and the Knut and Alice Wallenberg Foundation. Training of the neural network was performed using resources provided by the Swedish National Computing Infrastructure (NAISS) at Chalmers/C3SE and KTH/PDC, with part funding from the Swedish Research Council. Part of this work was carried out within Chalmers’ META-PIX Competence Center.
Never miss a breakthrough: join the SciTechDaily newsletter.
Follow us on Google and Google News.
