Researchers have used a neural network to identify twisted two-dimensional crystal configurations capable of steering light at the nanoscale, according to a report published in Nature Materials on August 28, 2026.
The neural network operates by working backwards from a chosen target light path. It calculates the specific multilayer crystal arrangements needed to generate the desired propagation across operating frequencies that span from the visible spectrum to terahertz ranges.
Light control in these systems occurs through surface polaritons generated in two-dimensional layers of van der Waals crystals stacked with relative rotational angles. The work builds on research authored by L. F. Álvarez-Tomillo and colleagues regarding deep learning design for nanoscale polariton propagation in twisted multilayers.
The underlying physics builds on earlier experimental investigations into anisotropic alpha-molybdenum trioxide (α-MoO3) crystals. Research published in 2020 demonstrated that rotating α-MoO3 layers reshapes phonon-polariton dispersion into directional light paths, while accompanying studies discovered photonic magic angles and diffraction-resistant canalized polariton modes in twisted bilayers. Later findings in 2023 established that three-layer α-MoO3 stacks provide additional rotational freedom, enabling multiple tunable and spectrally robust canalization regimes.
Springer Nature issued the peer-reviewed report under DOI 10.1038/s41563-026-02744-x. The publisher prices individual PDF access to the article at $39.95, with 30-day digital access offered through Nature+ for $32.99 and annual print subscriptions covering 12 issues set at $259.00.
