Physics-Informed Neural Networks Cut Optical Material Design Time From 30 Days to 3
Researchers at Chalmers University of Technology in Sweden embedded electromagnetic laws directly into a neural network's architecture, reducing the time needed to design optical materials from a month to three days. The method produces accurate predictions with a fraction of the training data conventional neural networks require.
The work addresses a real bottleneck. Designing materials that control light at nanoscale requires testing candidate structures in computer simulations, where each data point takes 10 minutes to an hour to generate. Standard neural networks demand massive training datasets before they can predict how a structure will scatter, reflect, transmit, or absorb light.
Building Physics Into the Model
Instead of forcing the network to discover the laws of electromagnetism from raw data, the team built those laws into the model itself. They embedded an analytical physics framework based on quasinormal modes-a method that connects scattering behavior to the natural resonances of optical devices-directly into the neural network's final stage.
"When we fed the super-brain information about the laws of physics, it immediately got much smarter," said Philippe Tassin, professor in the Department of Physics and Astronomy at Chalmers. "Our calculations now take one tenth of the time previously required."
The resulting model, called QNM-Net, learned physics-based parameters instead of raw simulation outcomes. This shift changed what the network needed to master.
Less Data, Better Results
Testing on photonic crystal slabs-patterned dielectric sheets whose scattering response centers on a single resonance-showed the advantage clearly. QNM-Net achieved low prediction error using about 160 training samples. Conventional neural networks needed roughly 10 times more data and more trainable parameters to reach similar accuracy.
The model did more than match prediction targets. Its learned resonance frequencies aligned closely with eigenfrequencies from full-wave eigenmode simulations, suggesting the internal physics it learned was physically meaningful, not just numerically useful.
On a harder test case-a free-form dielectric metasurface with multiple overlapping resonances and polarization-dependent behavior-QNM-Net reached the performance of the best reference models using about one-third as much training data.
Speeding Up Design Cycles
The speed gain matters for actual design work. Researchers used the learned resonance parameters as design targets to inverse-design photonic crystal slabs. The optimization reached desired eigenfrequencies after a few hundred steps in less than one second, with full-wave simulations later confirming the predictions.
Viktor Lilja, a doctoral student on the project, described the practical benefit: "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."
Why This Matters for Development Teams
The approach points to a broader shift in how machine learning integrates with physics-based work. Rather than treating neural networks as pattern-matching machines that need enormous datasets, researchers are baking scientific structure into models from the start.
For teams building optical components, photonic crystals, or metasurfaces, faster design cycles mean less time waiting for simulation data to accumulate. The method also proved more robust to simulated noise than standard neural networks, suggesting it could work with real-world experimental data where measurements are messy.
Because the network predicts interpretable physical parameters, it can also highlight which resonances matter most and which modes contribute little to observable scattering response. That transparency helps engineers understand why a design works, not just that it does.
The research was published in Laser & Photonics Reviews.
Your membership also unlocks:
