Polaritons are quasiparticles formed when electromagnetic radiation interacts strongly with an excitation in matter. In the systems examined here, the relevant excitations are optical phonons—collective vibrations of the crystal lattice. When infrared light couples to these vibrations, the resulting phonon polariton carries both light-like and matter-like properties. Because its electromagnetic field is concentrated near the material surface, a polariton can be squeezed into dimensions far smaller than the wavelength of the original light in free space. This confinement makes polaritons attractive for nanoscale imaging, sensing, information processing and thermal control, but it also makes their propagation highly sensitive to the crystal’s structure and orientation. In an ordinary isotropic material, energy tends to spread in a roughly circular wavefront. In an anisotropic material, however, the permitted relationship between frequency and momentum—the dispersion relation—can stretch or redirect the wave. The direction in which energy travels is governed by the group velocity, while the direction of the polariton wavevector describes the phase evolution. When these directions become strongly constrained, light can travel in narrow channels rather than diffracting broadly.

The researchers exploit this behavior by stacking and twisting van der Waals layers. These materials are built from atomically thin sheets held together by relatively weak forces between layers, allowing one sheet to be rotated with respect to another. That twist changes the combined electromagnetic environment and therefore the dispersion landscape experienced by the polaritons. In α-MoO₃, the crystal is naturally anisotropic: its response to infrared radiation differs along different in-plane crystallographic axes. A single layer can therefore support highly directional propagation, while a twisted multilayer can reshape and combine those directional responses. The resulting structure acts less like a passive slab and more like a nanoscale landscape in which certain propagation directions are favored. Previous experiments and theoretical work have shown that twisted van der Waals materials can produce canalization, a regime in which polariton energy travels predominantly along a narrow direction with minimal lateral spreading. The new study extends that concept by asking whether several distinct channels can be engineered deliberately, rather than discovered after a structure has already been selected.

To search this enormous design space, the team used deep neural networks. A multilayer polaritonic structure is defined by parameters such as the number of layers, their relative twist angles and the frequency at which the system is excited. Each combination can generate a different dispersion surface and a different real-space propagation pattern. Directly calculating every possible configuration is computationally expensive, especially when the desired outcome is specified first and the structure must be found afterward. A trained network provides a faster inverse-design route: it learns the relationship between structural parameters and polariton behavior, then uses that learned relationship to propose a configuration for a target outcome. In this case, the targets include canalization, in which propagation is concentrated into one dominant channel, as well as bicanalization and tricanalization, in which two or three preferential channels emerge. The significance is not simply that a machine-learning model can classify patterns. It is that the model is used as a design tool, linking an abstract optical requirement to a concrete twisted multilayer architecture.

The resulting predictions reveal canalization, bicanalization and tricanalization in twisted α-MoO₃ homostructures across the previously unexplored frequency range of 600–800 cm⁻¹. Wavenumber, measured in inverse centimetres, is commonly used for infrared spectroscopy and is related to frequency through the speed of light; the reported range lies in the mid-infrared portion of the spectrum. At these frequencies, the crystal’s optical phonons strongly influence the dielectric response, making α-MoO₃ capable of supporting hyperbolic phonon polaritons. In a hyperbolic medium, the in-plane components of the dielectric tensor have different signs, producing an open hyperbolic dispersion contour rather than the closed contours familiar from ordinary materials. This unusual geometry permits large wavevectors and highly compressed wavelengths. By rotating multiple layers, the researchers alter the orientation and intersection of the relevant dispersion features. The effect is to guide polariton energy into selected directions, potentially allowing a single nanoscale platform to split, merge or route energy through several channels. The study’s demonstrations show that the neural-network approach is not restricted to one especially favorable frequency or one isolated propagation effect.

One of the most practical tests involved an existing α-MoO₃ bilayer. Rather than designing an entirely new multilayer from the beginning, the researchers added an extra α-MoO₃ sheet to the structure and used the method to obtain a specified propagation behavior. This is an important distinction for experimental nanophotonics, where fabrication constraints can make a complete redesign costly or impossible. Two-dimensional and van der Waals materials can be assembled layer by layer, but each additional interface, alignment step and twist angle introduces potential sources of uncertainty. A design strategy that works with an established structure and identifies a minimal modification could make optical devices easier to tune. The result also illustrates the broader promise of inverse design: the desired function can be treated as the starting point, while the material geometry becomes the variable to be solved for. In principle, that could allow researchers to request a propagation pattern suited to a particular sensing geometry or thermal pathway, then use the model to determine how the layers should be arranged.

The team further extended its neural networks beyond α-MoO₃ to a variety of other materials. According to the study, the models can predict canalization from the visible to the terahertz regime, suggesting that the underlying strategy is not tied to one narrow spectral band. Different materials support different resonances, anisotropies and loss mechanisms, so transferring the method requires the network to account for each material’s electromagnetic response. The visible, infrared and terahertz regions also differ substantially in wavelength, available fabrication methods and sources of optical excitation. Nevertheless, the common design principle remains the same: calculate or learn how a material’s dispersion controls energy flow, then identify the structure that produces the desired directionality. Extending the predictions across this broad range could connect nanophotonic concepts that are often treated separately. A design platform capable of operating across multiple spectral regimes might help researchers adapt the same conceptual device architecture to different wavelengths, rather than developing a new trial-and-error process for every application.

The most immediate scientific value of the work is therefore methodological. Nanoscale polariton propagation depends on a complicated interplay among crystal anisotropy, layer orientation, interlayer coupling, frequency and wavevector. Small changes in any one of these variables can reshape the available optical modes. Conventional intuition remains essential, but it is poorly suited to exploring all combinations of parameters, particularly when several layers can be rotated independently. Deep learning can provide a way to navigate that high-dimensional landscape, while the physics supplies the constraints that keep the search connected to real materials. The approach does not make the underlying electromagnetic laws disappear; rather, it offers a rapid approximation to the relationship between a multilayer’s geometry and its polariton response. Its reliability will ultimately depend on how accurately the training data represent fabrication tolerances, material losses and imperfections. A predicted channel is useful only if it survives the deviations that occur when atomically thin layers are assembled in the laboratory. Even so, the demonstrated designs indicate that machine learning can move beyond recognizing optical patterns toward actively specifying them.

Programmable polariton propagation could have consequences well beyond the specific structures studied. In sensing, tightly confined fields can increase interactions with molecules or nearby surfaces, while directional channels could help deliver signals to a detector with reduced background. In thermal management, phonon polaritons can influence how electromagnetic energy moves through and away from nanoscale structures, potentially offering new ways to shape radiative heat flow. Directional propagation may also be relevant to infrared components that need to guide energy around compact circuits without relying on conventional dielectric waveguides. Canalization, in particular, suppresses the spreading that normally limits resolution and device density, while multiple channels could support more complex routing functions. The study does not report a finished sensor or thermal device, and the predictions will require experimental validation across the proposed material combinations. But by pairing twisted van der Waals engineering with an inverse-design neural network, it establishes a framework for turning polaritons from naturally occurring excitations into more deliberately arranged nanoscale carriers of energy. The broader vision is a form of artificial-intelligence-assisted nano-optics in which researchers specify how light should move first—and then construct the layered crystal that makes that motion possible.