Electromagnetic (EM) invisibility achieved via reflection and scattering suppression is becoming a key enabling function for optically transparent platforms, including transparent electronics, vehicle glazing, and smart-building windows. Unlike conventional low-observability coatings, transparent EM-stealth surfaces must satisfy a tightly coupled set of requirements: broadband low reflection over microwave-to-millimeter-wave frequency bands; a thin, lightweight form factor suitable for integration; high visible transmittance; and stable performance under oblique incidence and varying polarizations. Meeting these requirements simultaneously remains difficult. Traditional microwave absorbers are typically optically opaque, whereas transparent absorber designs must still contend with the thickness–bandwidth trade-off and the strong coupling among multilayer spacing, sheet resistance, and geometry. As a result, the design space is high-dimensional and highly nonconvex; brute-force sweeps or trial-and-error full-wave optimization quickly become impractical when wideband, wide-angle, and optical constraints are imposed simultaneously. To overcome these barriers, we propose an intelligent inverse-design framework that integrates deep learning with a genetic algorithm (GA), termed the deep genetic algorithm (DGA). The central strategy is to accelerate optimization without sacrificing physical fidelity by coupling (i) a data-driven forward surrogate model that rapidly predicts broadband EM responses with (ii) a global evolutionary search that can enforce multiple constraints and navigate complex, multimodal landscapes. We first generate a large-scale dataset through automated full-wave simulations (CST) driven by external scripting, covering key design degrees of freedom for transparent multilayer absorbers, including unit-cell periodicity, multilayer air-gap thicknesses, ring-pattern feature parameters, and the sheet resistances of multiple indium tin oxide (ITO) layers. A forward-prediction network (convolutional feature extraction followed by regression layers) is then trained to learn the nonlinear “structure-to-spectrum” mapping and to output the multi-frequency reflection response with millisecond-level inference latency. This surrogate does not replace physics; rather, it replaces repeated full-wave evaluations inside the optimization loop, thereby enabling efficient exploration of the design space. On this basis, DGA performs constrained inverse design using GA operations (selection, crossover, mutation, and elitism). The objective formulation targets ultra-wideband low reflection (equivalently, high absorption for a backed absorber with negligible transmission) while enforcing practical fabrication constraints, especially total thickness and angular robustness. To avoid designs that merely satisfy a narrowband criterion, we combine pointwise broadband reflection constraints with band-wise statistical constraints across standard sub-bands, which encourages consistently low reflection across the full target spectrum and improves robustness under oblique incidence. Using DGA, we design and fabricate an ITO-based, optically transparent absorbing metasurface composed of patterned ITO/PET films separated by air gaps and backed by a low-sheet-resistance ITO reflective layer. The optimized architecture employs three patterned ITO layers with differentiated sheet resistances and geometries to create distributed, coupled resonances that overlap across the frequency band, while the air gaps tune interlayer coupling and phase accumulation to realize broadband impedance matching. The unit cell exhibits rotational symmetry, supporting polarization-insensitive responses under both normal and oblique incidence. Experiments validate both the design methodology and the device performance. The fabricated sample maintains an average visible transmittance above 60%, confirming practical optical transparency. EM measurements demonstrate ultra-wideband absorption from 3.5 to 35 GHz (upper S band to Ka band) with a fractional bandwidth of 163% under the ≥90% absorption criterion, while maintaining stable, polarization-insensitive performance over incident angles from 0° to 50°. Minor deviations between simulation and measurement are attributable to manufacturing tolerances in ITO sheet resistance, assembly-induced spacing errors, and residual perturbations from the measurement support structure. Overall, DGA offers a scalable and practical pathway for designing optically transparent, ultra-wideband, and wide-angle EM-stealth metasurfaces under realistic constraints, and the demonstrated prototype supports integrated “optical transparency–EM stealth” functional surfaces for transparent electronics protection, stealth/EMC windows, and building-integrated EMC applications.
Zhao et al. (Thu,) studied this question.