Fig 1

Conceptual schematic of the sequence-aware inverse design framework, illustrating the workflow from impregnation experiments to the final optimal gradient structure."

Fig 2

Detailed computational architecture of the hybrid methodology. (a) Structure of the process-property database. (b–d) Internal architecture of the LSTM surrogate model. (e) Workflow of the Genetic Algorithm operators."

Fig 3

SEM images of (a1‒c1) pure PU foam, (a2‒c2) PU@CNT1, (a3‒c3) PU@CNT22, and (a4‒c4) PU@CNT333. (d) Schematic diagram of CNT coating. (e) EDS elemental analysis."

Fig 4

(a) XRD patterns and (b) Raman spectra of the pure PU foam and PU/CNT composites with different impregnation histories (PU@CNT1, PU@CNT22, and PU@CNT333). (c) FTIR spectrum of the pure PU foam skeleton."

Fig 5

3D waterfall plots of the (a) real part (ε′) and imaginary part (ε′′) of complex permittivity, (b) dielectric loss tangent (tanδε), and (c) RL for the 18 training dataset samples across the 2–18 GHz frequency range."

Fig 6

Performance evaluation of the LSTM surrogate model. Regression plots of predicted vs. experimental values for (a) real permittivity (ε′) and (b) imaginary permittivity (ε′′). Comparison of R2 and RMSE metrics for the training and test sets for (c) ε′ and (d) ε′′."

Fig 7

SHAP interpretability analysis for the real part of permittivity (ε′). (a) Feature importance bar chart. (b) SHAP bee swarm plot. (c) Dependence plot for Step2 concentration. (d) Interaction dependence plot for Step2 vs. Frequency. (e) Dependence plot for Frequency. (f) Interaction dependence plot for Frequency vs. Step2."

Fig 8

Evolutionary optimization process. (a) Schematic of the three-layer gradient absorbing structure model. (b) Fitness evolution curve over 20 generations. (c) RL curves and (d) 3D permittivity waterfall plots for the best individuals from the 1st, 2nd, 3rd, and 20th generations."

Fig 9

Experimental validation and loss mechanism. (a) Complex permittivity distribution of the optimized three-layer gradient foam. (b) Photograph of the MA measurement setup using the arch method. (c) Comparison between the experimentally measured and LSTM-predicted RL curves. (d) Simulated power loss density distributions inside the gradient foam at 4, 10, and 16 GHz under TE and TM polarizations."

Fig 10

Radar Cross Section (RCS) reduction performance. (a) Schematic illustration of stealth application. (b) Monostatic RCS polar plots of PEC and OGS-Foam at 4, 10, and 16 GHz under TM and TE polarizations."