Fig. 1. Schematic of this work: a multispectral compatible metasurface is designed via machine learning, which is compatible with VIS, infrared, and microwave frequencies.

Fig. 2. Structure design: (a) three-dimensional structure diagram and geometrical parameter of the unit, (b) the architecture of BPNN, and (c)–(e) details of neurons.

Fig. 3. Process and results of NN training: (a) the variation of MSE in the training process, (b) examination of the training process, (c) error histogram of trained network, and (d) the regression plot of all the samples.

Fig. 4. Prediction and verification of structures and EM response by the NN: (a) six generated curves, (b) structural parameters of network prediction, (c) simulation results of the predicted structures, (d) the structures corresponding to the predicted parameters, (e) structural parameters verified by CST program for different materials, and (f) comparison between the input curves and the predicted curves with different materials.

Fig. 5. Structure optimization and mechanism analysis: (a) three-dimensional structure diagram and geometrical parameter of the optimized structure, (b) the simulation of co-polarization reflectivity (Ryy) at different incident angles, (c) the efficiency of different mechanisms including co-polarized reflectivity (Ryy), cross-polarized reflectivity (Rxy), and absorptivity (A), (d) the efficiency proportion of different mechanisms, (e) the E-field distribution, and (f) the surface current distribution.

Fig. 6. Sample fabrication and performance characterization: (a) the processing of the metafilm, (b) the details of the fabricated metafilm, (c) the measurement of VIS light transmittance, and (d) demonstration of the flexible metafilm.

Fig. 7. Measurement environment and results: (a) microwave measurement environment, (b) the measured and simulated reflectivities of the sample, (c) the measurement of mean infrared emissivity, and (d) the measurement of infrared emissivity at 3–14 μm.

Fig. 8. Measurement of infrared radiation performance: (a) ITO (10 Ω/sq), PMMA, PET, and the metafilm on heating platform, and (b)–(d) infrared radiation detection at 28°C, 50°C, and 100°C.

Fig. 9. Euclidean distance matrices: (a) and (c) distance matrices of reflectivity curves, and (b) and (d) distance matrices of structural parameters.

Fig. 10. Performance of BPNN for different materials: (a)–(c) the training states of BPNN for F4B, FR4, and PI materials and (d)–(f) the error of trained BPNN for F4B, FR4, and PI.

Table 1. Training Performance of Different Dielectric Substrates