At Chalmers University of Technology in Gothenburg, Sweden, physicist Philippe Tassin made a discovery that feels almost counterintuitive: teaching a machine learning system the laws of physics made it smarter, not constrained. His team has developed a physics-informed neural network that accelerates the design of nanophotonic components—the tiny optical materials that could make future quantum computers, camera lenses, and eyeglasses lighter, thinner, and far more effective—by a factor of ten.
The challenge they faced was fundamentally one of time. Designing advanced optical materials requires running thousands of computer simulations, each one checking how light behaves in specially engineered structures smaller than the wavelength of light itself. Generating a single data point for training their neural network could take anywhere from ten minutes to an hour. To properly train the system, researchers like doctoral student Viktor Lilja might need 40,000 simulations—work that stretched across entire months.
Then came the insight that changed everything. Rather than forcing the neural network to rediscover the laws of electromagnetism and physics on its own through trial and error, Tassin's team did something radical: they taught the neural network those laws directly. They embedded fundamental physics equations into the system before training began, essentially giving the artificial intelligence a physics education before asking it to solve problems.
The results were striking. What previously took thirty days to generate now takes three days. The neural network needed far less training data to reach the same level of accuracy because it no longer had to "reinvent the wheel," as Tassin puts it, by deriving basic physical principles from raw data. "When we fed the super-brain information about the laws of physics, it immediately got much smarter," Tassin said. "Our calculations now take one tenth of the time previously required."
The practical implications ripple outward quickly. In the near term, this approach could accelerate the development of optical materials that transform everyday technology—making eyeglass and camera lenses more efficient and compact. But the vision extends further into emerging frontiers. Tassin's group is collaborating with researchers at Chalmers' Department of Microtechnology and Nanoscience, where Sweden's first larger quantum computer is being built, to explore whether nanostructured photonic crystals could help transmit quantum information over longer distances using light itself. These man-made crystals have an extraordinary capacity to reflect and control light, and optimizing their design through simulation could be key to scaling up quantum networks.
What makes this breakthrough particularly elegant is that it emerged almost accidentally. Tassin and his team weren't initially trying to speed up their calculations; they were simply attempting to make the neural network's predictions easier for humans to interpret by grounding them in recognizable equations. Only when they tested the system did they discover the unexpected payoff—faster learning, fewer data requirements, and demonstrably more powerful predictions.
The research, published in Laser & Photonics Reviews, represents a subtle but significant shift in how artificial intelligence tackles scientific problems. Rather than treating neural networks as pure black boxes, this approach treats them as tools that can be enhanced by the knowledge humans have already accumulated. Physics still matters. It just needed a smarter way to teach it.