AI model designs complex metasurfaces in seconds using physics-guided learning, enabling rapid, high-quality holographic devices and adaptable optical systems for multiple wavelengths and polarizations.
(Nanowerk Spotlight) Imagine a holographic display as thin as a sheet of glass, projecting crisp, full-color 3D images without lenses or bulky optics. The same flat device could focus light for medical imaging, filter specific wavelengths for scientific instruments, or conceal information in patterns invisible to the naked eye. These capabilities depend on metasurfaces — ultrathin materials patterned with millions of nanostructures that bend and shape light in ways conventional optics cannot.
The promise is clear, but the obstacle is design speed. Each nanostructure, called a meta-atom, must be precisely tuned to control properties such as brightness, color, polarization, and phase. Standard “inverse design” methods work backwards from the desired output, running thousands of simulations and adjusting each structure until the result matches the target. For high-resolution or multi-functional devices, this can take hours or days. Any change to the image or optical function forces the entire process to start over. Even newer AI-assisted approaches, which replace some simulations with trained predictions, still require full-surface iterative optimization, leaving the bottleneck intact.
Recent advances in computing hardware, neural network architectures, and physics-informed learning are making a different strategy possible. In other areas of photonics and imaging, researchers have shown that deep learning can map directly from a desired optical output to the structure needed to create it. Self-supervised learning, which uses physical models instead of labelled datasets, combines accuracy with efficiency and avoids the heavy cost of data preparation. Together, these developments suggest metasurface design could shift from a long trial-and-error process to a one-step prediction.
That shift is exactly what researchers in China aim to achieve with their physics-driven self-supervised network (PDSS-Net). Using meta-holography as a demanding test case, the system learns a direct mapping from a target image to the structural parameters of every meta-atom. Once trained, it can generate complete high-resolution designs in under a second, with results that match or exceed the quality of slower, iterative methods.
Schematic of the multidimensional multiplexed meta-holography enabled by the PDSSNet. Various images with different polarizations and colors on two imaging planes can be selectively concealed or revealed by switching the output polarization states, where each meta-atom simultaneously controls the amplitude, phase, wavelength, and polarization of the light field on demand. (Image: Reprinted from DOI:10.1002/advs.202509242, CC BY) (click on image to enlarge)
The key advantage of PDSS-Net is that it removes the main computational bottleneck by skipping the element-by-element optimization loop. It treats design as a direct translation problem — an image in, a fabrication-ready pattern out. This approach also adapts readily to new functions. By retraining with different datasets, the same network architecture can handle new combinations of wavelengths, polarizations, or depths without structural changes.
PDSS-Net combines three components to achieve this. First is an encoder–decoder network that takes a multi-channel target image as input, with each channel representing a different optical property to be controlled. The encoder compresses the input into a set of features describing the design requirements, and the decoder reconstructs these into a map of each meta-atom’s length and width. These dimensions determine how the structures scatter light, while height and spacing are fixed for fabrication consistency. To make the designs robust, the network introduces small random variations to these dimensions during training, simulating the imperfections that occur during manufacturing.
The second component is a deep neural network trained to predict how an individual meta-atom will affect light. This prediction is expressed as a Jones matrix, a standard mathematical tool for describing changes in amplitude and phase for different polarizations. The network takes the length and width of a meta-atom as inputs and outputs the Jones matrix values for the wavelengths of interest, replacing slow electromagnetic simulations with instant predictions.
The third component is a physics-based module that uses these predicted values to calculate the full metasurface output. It applies the angular spectrum method, which models how light propagates after passing through the surface. Because the output can be compared directly to the target image, the system can train itself without manually prepared structure–image pairs. This makes the method self-supervised and ensures that the designs follow the laws of physics.
Training is guided by a composite loss function combining several measures of image fidelity: Pearson correlation for similarity, structural similarity for texture and pattern accuracy, and peak signal-to-noise ratio for clarity. Once trained, PDSS-Net can design a metasurface for a new target in a single step.
The researchers first tested the method on three-wavelength, 2K-resolution meta-holograms. Each wavelength corresponded to a separate color channel in the reconstructed image. PDSS-Net produced a complete design in 0.8 seconds on a commercial workstation. An AI-assisted iterative method took over 1700 seconds. The PDSS-Net results also had higher measured image quality, which the authors link to its ability to integrate spatial context rather than optimizing each element independently.
Retraining the network allowed the team to tackle more complex designs. One example was a 16-channel scalar hologram, controlling four wavelengths, two polarizations, and two image depths independently. Each channel displayed a different character, and the network produced the designs in under half a second. When fabricated, the devices showed low crosstalk between channels, even for closely spaced wavelengths. Because the network had been trained with simulated fabrication errors, performance remained high even when actual errors exceeded 12 nanometers.
The team also demonstrated vectorial meta-holography, which adds control over spatially varying, non-orthogonal polarization patterns alongside brightness and phase. This enables displays that change appearance or reveal hidden details when viewed through specific polarization filters. In their demonstration, the metasurface generated two full-color images at different depths, each with its own complex polarization pattern. These patterns determined which parts of the image were visible under different analysis polarizations. The designs required a brief fine-tuning step for each target, adding about 40 seconds to the process, but still achieved a large speed advantage over traditional methods.
Although this study focuses on holography, the same framework could be applied to other metasurface devices such as flat lenses, beam shapers, or polarization-sensitive imagers by changing the training data and optimization criteria. The authors suggest that integrating geometric phase techniques, expanding the diversity of training datasets, or adopting transformer-based architectures could further enhance performance. Transfer learning could also allow the network to adapt quickly to new tasks, and the approach could be extended to metasurfaces made from tunable materials such as liquid crystals or phase-change compounds.
The broader achievement is moving metasurface design from a slow, iterative process to a rapid, physics-guided prediction. This capability enables the creation of high-quality designs in seconds, opening the door to large-scale manufacturing and real-time customization in applications where quick turnaround is essential. With ongoing advances in fabrication and computing, methods like PDSS-Net could help metasurfaces transition from laboratory prototypes to widely used components in displays, imaging systems, communications, and optical security.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
