Abstract
Topological photonic crystals (PhCs) that support disorder-resistant modes, protected degeneracies, and robust transport have recently been explored for applications in waveguiding, optical isolation, light trapping, and lasing. However, designing PhCs with prescribed topological properties remains challenging because of the highly nonlinear mapping from the continuous real-space design of PhCs to the discrete output space of band topology. Here, we introduce a machine learning approach to address this problem, employing Kolmogorov--Arnold networks (KANs) to predict and inversely design the band symmetries of two-dimensional PhCs with two-fold rotational (C2) symmetry. We show that a single-hidden-layer KAN, trained on a dataset of C2-symmetric unit cells, achieves high accuracy in classifying the topological classes of the lowest lying bands. We use the symbolic regression capabilities of KANs to extract algebraic formulas that express the topological classes directly in terms of the Fourier components of the dielectric function. These formulas not only retain the full predictive power of the network but also provide novel insights and enable deterministic inverse design. Using this approach, we generate photonic crystals with target topological bands, achieving high accuracy even for high-contrast, experimentally realizable structures beyond the training domain.
The Challenge of Topological Design
Topological photonic crystals (PhCs) represent a frontier in modern optics, promising light transport immune to disorder and defects. However, a significant "mapping gap" hinders their development: designers must bridge the gap between continuous real-space material arrangements and discrete, integer-valued topological invariants. Traditional inverse design often relies on "black-box" optimization or exhaustive searches, which offer little physical intuition and struggle to generalize across different material systems.
Interpretable Machine Learning via KANs
To overcome these limitations, this research moves away from standard neural networks in favor of Kolmogorov-Arnold Networks (KANs). Unlike standard Multi-Layer Perceptrons (MLPs) that use fixed activation functions on nodes, KANs employ learnable 1D splines on the edges. This architecture is exceptionally efficient at identifying sparse, interpretable patterns within complex data.
A key discovery made through KANs is the "Rule of Three." Despite being trained on 19 input Fourier parameters, the network automatically pruned itself to just three critical coefficients: $\epsilon_{01}$, $\epsilon_{10}$, and $\epsilon_{11}$. This reveals that the topology of the lowest bands is governed almost entirely by the coarsest geometric features of the crystal. By applying symbolic regression, the researchers distilled the trained network into eight explicit algebraic formulas that link these Fourier components directly to topological classes, effectively "rediscovering" the underlying physics of degenerate perturbation theory.
Deterministic Inverse Design and Generalization
The extracted symbolic formulas enable a deterministic inverse design workflow. By selecting a target topological class, researchers can use these formulas to perform gradient descent and find the exact Fourier coefficients required for a specific design.
Remarkably, although the model was trained on "smooth," low-contrast dielectric profiles, the resulting formulas maintained over 92% accuracy when applied to "two-tone" (binary) high-contrast structures, such as silicon pillars in air. This generalization is critical for experimental realization. Furthermore, the physics-informed data augmentation—shifting unit cell centers to different Wyckoff positions—allowed the model to achieve ~90% accuracy with as few as 64 training samples.
Conclusion and Future Outlook
This "symbolic learning" approach successfully transforms machine learning from a black box into a transparent tool for physical discovery. The researchers demonstrated its power by designing crystals with frequency-isolated Dirac points and robust edge states using only their symbolic formulas as a guide. While currently focused on 2D $C_2$ symmetry, this methodology could be extended to higher symmetries and 3D structures, potentially revolutionizing materials science by predicting complex electronic or optical properties directly from simple density maps.
Summary updated