- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,246
The neural network was taught to understand the physics of light even before training, and this approach sharply accelerated the design of optical materials. Researchers from Chalmers University of Technology in Sweden have built into the model the basic laws of electromagnetism, after which the calculations began to take ten times less time. The development can accelerate the creation of new lenses, photonic crystals and components for quantum technologies.
The team works in the field of nanophotonics, where scientists control light at scales less than wavelength. In such conditions, light behaves not as in conventional optics, and engineers get the opportunity to create artificial materials with properties that are not in nature.
With the help of supercomputer simulations, the researchers select structures for materials that can make camera lenses and eyeglass thinner, lighter and more efficient. The same technology can be useful in quantum systems where it is necessary to accurately control the propagation of light.
Together with colleagues from the Department of Microtechnologies and Nanoscience, where Sweden’s first major quantum computer is being built, the group is studying nanostructured materials for transmitting information between quantum computers or over long distances using optical frequencies. In one direction, scientists are looking at mechanically malleable photonic crystals that can direct light in the right way.
The main problem was the speed of calculations. Neural networks can analyze huge arrays of simulations and predict how the material behaves, but to train such systems need a lot of data. According to the researchers, the calculation of one data point took from ten minutes to an hour, and the full set could require up to 40 000 simulations.
“I know the equations of electromagnetism along and across and teaching them, but still I can’t draw all the conclusions that the neural network makes. Physics is so complex that I do not understand the properties of the material simply by appearance, and the computer understands,” said Professor of the Department of Physics and Astronomy Philip Tassen.
To remove a bottleneck, scientists did not force the model to rediscover the laws of electromagnetism according to data. Instead, the researchers immediately built a basic understanding of the behavior of light and electromagnetic fields into the neural network.
Initially, the team wanted to draw the conclusions of the neural network more understandable to people, adding to the model of equations familiar to physicists. During the tests, it turned out that this approach not only simplifies the interpretation, but also dramatically increases the efficiency of calculations.
“When the network is already trained, we can ask it to check almost any structure and get optical properties in a millisecond. With the new networks, we get more accurate estimates and avoid obvious mistakes,” said researcher Victor Lilya.
Philip Tassen compared the new approach to the “supermozed”, which was explained in advance by physical rules. “When we gave the superbrain information about the laws of physics, it immediately became much smarter. Now our calculations occupy one-tenth of the previous time,” said the professor.
As a result, the time of preparation of data for simulations was reduced from 30 to three days. For the development of optical components, this difference is of practical importance: engineers can quickly test new structures, weed out unsuccessful options and search for materials for the next generation of lenses, photonic devices and quantum systems.
The new approach helps to create nanophoton materials for cameras, glasses and quantum systems faster.
The team works in the field of nanophotonics, where scientists control light at scales less than wavelength. In such conditions, light behaves not as in conventional optics, and engineers get the opportunity to create artificial materials with properties that are not in nature.
With the help of supercomputer simulations, the researchers select structures for materials that can make camera lenses and eyeglass thinner, lighter and more efficient. The same technology can be useful in quantum systems where it is necessary to accurately control the propagation of light.
Together with colleagues from the Department of Microtechnologies and Nanoscience, where Sweden’s first major quantum computer is being built, the group is studying nanostructured materials for transmitting information between quantum computers or over long distances using optical frequencies. In one direction, scientists are looking at mechanically malleable photonic crystals that can direct light in the right way.
The main problem was the speed of calculations. Neural networks can analyze huge arrays of simulations and predict how the material behaves, but to train such systems need a lot of data. According to the researchers, the calculation of one data point took from ten minutes to an hour, and the full set could require up to 40 000 simulations.
“I know the equations of electromagnetism along and across and teaching them, but still I can’t draw all the conclusions that the neural network makes. Physics is so complex that I do not understand the properties of the material simply by appearance, and the computer understands,” said Professor of the Department of Physics and Astronomy Philip Tassen.
To remove a bottleneck, scientists did not force the model to rediscover the laws of electromagnetism according to data. Instead, the researchers immediately built a basic understanding of the behavior of light and electromagnetic fields into the neural network.
Initially, the team wanted to draw the conclusions of the neural network more understandable to people, adding to the model of equations familiar to physicists. During the tests, it turned out that this approach not only simplifies the interpretation, but also dramatically increases the efficiency of calculations.
“When the network is already trained, we can ask it to check almost any structure and get optical properties in a millisecond. With the new networks, we get more accurate estimates and avoid obvious mistakes,” said researcher Victor Lilya.
Philip Tassen compared the new approach to the “supermozed”, which was explained in advance by physical rules. “When we gave the superbrain information about the laws of physics, it immediately became much smarter. Now our calculations occupy one-tenth of the previous time,” said the professor.
As a result, the time of preparation of data for simulations was reduced from 30 to three days. For the development of optical components, this difference is of practical importance: engineers can quickly test new structures, weed out unsuccessful options and search for materials for the next generation of lenses, photonic devices and quantum systems.
The new approach helps to create nanophoton materials for cameras, glasses and quantum systems faster.
THE UNDERGROUND ARCHIVES 
Classified Files
Encrypted Communications
Dark Storylines
