Inverse design for laser-compatible infrared camouflage metasurface enabled by physics-driven neural network and genetic algorithm.

Metasurface-based multi-band camouflage holds significant value in both military and civilian applications due to its superior performance and structural simplicity. Combining a physics-driven neural network (PNN) and genetic algorithm (GA), we designed the AZO-Ge disk metasurface capable of achieving 1.06 μm laser stealth, mid-infrared thermal camouflage and efficient thermal management. In contrast to conventional spectra prediction neural networks, the PNN adopts a data dimensionality reduction prediction paradigm, resulting in enhanced accuracy and accelerated convergence. The combination of PNN and GA in the inverse design effectively addresses the prevalent "one-to-many" and single-solution challenges encountered in tandem neural networks. Such data-driven approach makes metasurface design faster and more efficient. Finally, four groups of metasurface geometric parameters were designed by the PNN-GA framework. The designed metasurfaces exhibit low average emissivities (<0.3 and < 0.2, respectively) in the two atmospheric windows (3∼5 μm and 8∼14 μm), high average emissivity (>0.75) in the non-atmospheric window (5∼8 μm) and high absorptivity (>0.8) at 1.06 μm. We also demonstrate that these spectral properties are insensitive to the incident angles and polarizations. In addition, the electric field intensity distribution plots of the cross sections reveal that surface plasmon polariton (SPP) predominantly contributes to the high emission in the non-atmospheric window band, while the high absorptivity at 1.06 μm is attributed to the resonance and the intrinsic absorption of the Ge material. The designed metasurface with its simple structure suggests the potential for low-cost and large-scale fabrication in the future. Meanwhile, the combination of machine learning and evolutionary algorithms is a pioneering work in metasurface design and we believe our work will provide guidelines for the multi-solution design of complex metasurfaces. • A framework that combine machine learning and genetic algorithm is adopted to design metasurface. • A neural network with physics-based data dimensionality reduction prediction. • An inverse design framework with multi-solution capability. • The generated solutions achieve laser stealth, thermal camouflage, and efficient thermal management. [ABSTRACT FROM AUTHOR]