Graph neural networks for structure-property prediction in photonic crystals
Аннотация
Photonic crystals (PCs) constitute a class of regular photonic materials characterized by a period comparable to the wavelength of light. Due to their ability to control and manipulate the motion of photons, PCs have been used to develop next-generation optical devices—including waveguides, resonators, amplifiers, splitters, and anti-counterfeit tags. The realization of such devices relies on PCs with designed photonic bandgaps or photonic band structures (PBSs), properties strongly connected to the lattice type and basis of the structures. The inverse design of PCs by identifying structure-property relationships is thus critical for practical engineering applications. Traditional machine-learning methods, such as regressions and linear discriminant analysis, are not well suited to extract structure-property relationships from complicated, high-dimensional data; others, such as decision trees and support vector machines, require explicit enumeration of structural features (e.g., volumes or aspect ratios of unit cells). Image-based models provide general methods for arbitrary structures but often demand extensive data sets and computational resources to capture complex relationships accurately. Moreover, the convolutional kernels in standard image models process uniform data structures and therefore may not fully address the nonuniformity inherent in PC unit cells. Graph neural networks (GNNs) provide a versatile approach to characterize and extract global features from space-variant media while remaining tensor-computable. In addition to photonic properties, PCs also support elastic waves, topological edge states, and other classical and quantum effects. The discovery of novel phenomena in periodic structures thus relies heavily on understanding PC spy graphs describing structure-property relationships.