The rapid expansion of modern wireless communication systems, including 5G base stations, satellite networks, and military radars, has dramatically intensified electromagnetic (EM) interference and radiation safety concerns in recent years [1], [2]. Conventional EM absorbers such as Salisbury screens, Jaumann structures, and honeycomb magnetic composites have been widely applied to suppress unwanted EM radiation. Salisbury screens employ a resistive sheet placed a quarter wavelength from a conductive backplane to achieve destructive interference and narrowband absorption [3], while multilayer Jaumann absorbers extend operational bandwidth through stacked resistive and dielectric layers at the expense of increased thickness and weight [4]. Additionally, magnetic composite absorbers utilizing ferrite or carbonyl iron powders embedded within polymer foams or honeycomb matrices offer relatively lightweight, broadband absorption in microwave regimes [5], [6], [7]. However, these structures are often constrained by limited angular stability and relatively narrow effective absorption bandwidths.

The emergence of metamaterials (MMs)—engineered subwavelength periodic structures capable of manipulating electric and magnetic resonances—has transformed absorber design methodologies [8]. Following the seminal experimental demonstration of negative refractive index metamaterials by Shelby [9], a variety of metamaterial absorbers (MMAs) have been proposed, offering tailored impedance matching and enhanced absorption across microwave to terahertz frequency ranges. Landmark work by Landy [10] introduced a “perfect” metamaterial absorber by coupling electric and magnetic resonances within a sandwich-type configuration. Subsequent developments included multilayer stacked MMAs [11], resistor-integrated metastructures [12], and coding metamaterials employing discrete pattern sequences for broadband EM wave manipulation [13], [14]. While these designs have substantially improved absorption efficiency and tunability, the realization of ultra-wideband (UWB) absorbers capable of sustaining high-performance absorption over frequencies exceeding 20 GHz remains technically demanding.

Notably, several persistent challenges hinder the advancement of UWB MMA design. One significant issue is the inherently many-to-one mapping between absorber structure and its EM absorption spectrum. Different configurations can produce nearly identical absorption responses, rendering inverse design non-unique and prone to solution ambiguities [15], [16]. Furthermore, absorber architectures typically involve high-dimensional, tightly coupled parameters – including dielectric thickness, surface resistance, unit cell geometry, and discrete coding patterns – whose complex interdependencies severely complicate manual design and render conventional heuristic optimization impractical for large-scale problems [13], [17], [18]. Achieving broadband impedance matching over extended frequency bands (e.g., 1.8–30 GHz) represents another formidable challenge, especially in coding-type absorbers where both discrete coding sequences and continuous physical parameters must be jointly optimized [19], [20]. Additionally, existing inverse design approaches often lack dedicated frameworks for coding MMAs and struggle to simultaneously balance computational efficiency, prediction accuracy, and scalability [21], [22], [23].

To address these limitations, recent research has explored data-driven machine learning (ML) strategies for EM absorber design and optimization. Convolutional neural networks (CNNs) have been widely applied to approximate the highly nonlinear mapping between structural parameters and absorption spectra, accelerating forward prediction and surrogate-assisted optimization [24], [25]. Variational autoencoders (VAEs) and tandem neural network (TNN) architectures have demonstrated capability in capturing latent features and facilitating inverse design for complex metastructures [21], [22]. Hybrid frameworks combining heuristic algorithms with deep learning models – such as CNN–GA (genetic algorithm) and particle swarm optimization-assisted neural networks – have also been developed to enhance global optimization efficiency and robustness [8], [23]. Despite these advances, comprehensive reviews by Qiu [15] and Lee [16] confirm that most existing methodologies remain constrained by high computational costs, restricted applicability to coding metamaterials, and unresolved multi-solution ambiguities in practical inverse design tasks.

In this study, we propose a novel double-layer indium tin oxide (ITO) coding metamaterial absorber featuring five tunable design parameters: dielectric thickness, top and bottom resistive film sheet resistance, periodicity, and coding pattern configuration. Two complementary inverse design frameworks are developed: (1) a CNN–GA hybrid strategy achieving starting and ending frequency deviation accuracies of 0.382/0.456 GHz, and (2) a TNN achieving a 17 fold acceleration in inverse design efficiency while maintaining >90% absorption in the 1.8–30 GHz band. Full wave electromagnetic simulations and equivalent circuit modeling further elucidate the absorption mechanisms, revealing broadband performance enhancement via collaborative electric and magnetic resonance coupling, Joule loss, and Fabry–Perot interference effects [5], [11]. The proposed framework offers a scalable, efficient, and interpretable data-driven solution for high-dimensional, ultra-wideband coding metamaterial absorber design.