The Future of Mechanical Metastructures: There is Plenty of Room at the Top

Abstract

Lightweight multi-functional mechanical metastructures (LM³) are architected structural systems whose performance is governed not only by constituent materials but also by topology, hierarchy, connectivity, and spatial organization across multiple length scales. Compared with conventional lightweight structures, LM³ provide opportunities to achieve combinations of load-bearing capability, energy absorption, vibration regulation, thermal management, sensing, and adaptive response that are difficult to realize through material selection alone. Recent advances in additive manufacturing, topology optimization, multi-scale modeling, and artificial intelligence (AI) have significantly expanded the design space of LM³ and accelerated their development toward engineering applications. As a timely response, this review summarizes recent progress in LM³ from the perspectives of structural design, manufacturing, performance regulation, and intelligent development. Particular attention is given to data-driven modeling, machine-learning-assisted prediction, generative and inverse design, bioinspired architectures, and AI-enabled optimization strategies. Emerging directions, including digital-twin-assisted topology tailoring, Structural Genome Initiative (SGI)-based design frameworks, high-throughput simulation and autonomous experimentation, large-language-model-driven multi-agent systems, and mechanoengineering-oriented methodologies, are also discussed. Finally, future challenges associated with structural genome representation, multimodal data integration, intelligent manufacturing, defect-aware validation, and service-oriented deployment are highlighted.

Graphical abstract

Keywords

Mechanical metastructures / Artificial intelligence / Structural Genome Initiative / Multi-agent systems / Mechanoengineering

Highlight

	● Reviews recent progress in lightweight multi-functional mechanical metastructures from mechanics, design, fabrication, and intelligent development perspectives.
	● Summarizes how structural genomes, hierarchy, topology, and manufacturable architectures enable property combinations beyond conventional material-property maps.
	● Discusses data-driven, physics-informed, generative, and inverse-design methods for mechanical metastructure prediction and optimization.
	● Proposes an SGI-based high-throughput framework linking multimodal databases, intelligent computation, automated experiments, and fabrication feedback.
	● Introduces mechanoengineering and AI4MechanoStructure as emerging frameworks for service-oriented mechanical metastructure design.

1 Introduction

1.1 Background, definition, and scope of lightweight multi-functional mechanical metastructures

In recent years, the rapid development of aerospace, intelligent manufacturing, and related industries, together with global efforts to reduce carbon emissions, has driven an urgent demand for lightweight, load-bearing, and multi-functionally integrated structures. Such structures are increasingly required in advanced industrial equipment, aerospace systems, and defense platforms that must operate under extreme service conditions and coupled multiphysics environments. Meanwhile, advances in computational mechanics, intelligent manufacturing, data science, and artificial intelligence have significantly expanded the capability for designing, optimizing, and manufacturing complex structural systems [1,2].

Mechanical metastructures can be understood as engineered structural systems whose effective mechanical properties and functional responses are determined primarily by architecture, including unit-cell topology, connectivity, hierarchy, and spatial arrangement, rather than by the intrinsic properties of the constituent materials alone [3,4]. This concept is closely related to mechanical metamaterials, but the emphasis is different. Mechanical metamaterials are often discussed from the perspective of unusual effective properties generated by artificial unit cells, whereas mechanical metastructures place greater weight on component-level integration, manufacturing feasibility, load-bearing function, and service performance [5,6]. This distinction reflects the different emphasis between scientific exploration and engineering deployment. In practical applications, manufacturability, structural integrity, reliability, and service performance at component and system scales are often as important as the effective properties achieved at the unit-cell level.

The aerospace sector offers a clear illustration of this need. Under Earth’s strong gravity, the overall transport efficiency of present launch vehicles is still quite low: the payload mass that ultimately reaches orbit is usually only about 1%–5% of the lift-off mass [7]. Even modest reductions in structural mass can significantly improve payload efficiency and mission economics [8]. Representative launch vehicles, such as Falcon 9 and Long March 5, also show that payload fractions for low Earth orbit are typically only several percent, with even lower values for higher-energy transfer orbits [9,10]. These figures highlight the continuing importance of lightweight structural design in aerospace engineering. However, reducing mass by replacing one material with a lighter one is no longer sufficient for many advanced applications. The structural system must also remain manufacturable, damage tolerant, multi-functional, and reliable over long periods of service. This requirement naturally shifts attention from material substitution alone to architecture-based structural design, in which topology, geometry, material, process, and performance are considered together.

This shift also exposes several unresolved challenges. The performance of mechanical metastructures can be strongly affected by nonlinear deformation, unit-cell connectivity, geometric scale effects, and the interaction between local architecture and the global structural response [11,12]. In practical manufacturing, especially additive manufacturing, geometric deviations, surface roughness, incomplete bonding, residual stress, and internal defects can produce measurable discrepancies between idealized designs and measured performance [13,14]. The resulting design problem is therefore high-dimensional: topology, geometry, material choice, process parameters, defects, loading history, and service environment may all influence the final response. These factors substantially increase design complexity and limit the efficiency of conventional trial-and-error approaches, creating a strong demand for data-assisted and physics-informed design methodologies. Accordingly, this review examines emerging data-driven and intelligence-enabled approaches for the design, manufacturing, and lifecycle optimization of lightweight multi-functional mechanical metastructures, with particular attention to structure–process–performance relationships and closed-loop design frameworks [15,16].

1.2 Development trends in lightweight materials and structures

Prior to World War Ⅱ, steel was the dominant industrial structural material and formed the backbone of modern industry [7]. In the postwar period, the rapid development of lightweight alloy systems and composite technologies fundamentally reshaped the structural materials landscape [17]. Aluminum alloys, titanium alloys, and fiber-reinforced composites, owing to their low density, high specific strength, and high specific stiffness, have been increasingly adopted in aviation and aerospace applications [10]. Advances in lightweight materials have also been accompanied by substantial progress in lightweight structural design methodologies, including mechanical design for composites and complex structures, lattice architecture design, structural topology optimization, bio-inspired design, and more recently, generative intelligent design [1,2,18,19]. This evolution reflects a broader transition from material-centered lightweight design toward integrated optimization of materials, architectures, manufacturing constraints, and service performance.

In general, the design of mechanical metastructures focuses on the topology of the unit cell, including its shape and connectivity, and on geometric parameters, such as angles, thicknesses, and curvature, to achieve lightweight and multi-functional performance [3,4]. This design concept is distinct from conventional topology optimization. In a standard topology optimization problem, the main goal is typically to make a given structure more efficient in terms of stiffness, weight, or related measures. By contrast, mechanical metastructures are designed not only to produce responses that are difficult to realize in conventional structural forms, but also to extend the performance limits summarized, for example, in the Gibson–Ashby charts for material properties [20]. The essential distinction is that mechanical metastructures treat architecture as a primary source of effective mechanical behavior, whereas conventional topology optimization usually redistributes material within a prescribed design domain [19].

Topology optimization provides a systematic framework for distributing material within a prescribed design domain to achieve specified performance objectives. The design domain is discretized into numerous finite elements, and local material distributions are iteratively updated according to optimization criteria. Through global coordination of these local design variables, the overall structural performance gradually converges toward the target objective [2]. Porous and architected lightweight materials and structures can be broadly classified into four main categories: stochastic porous foams, regular honeycombs and lattices, hybrid porous architectures, and metastructures [21-23].

In recent years, lightweight multi-functional mechanical metastructures (LM³) have attracted growing attention because of their broad application potential in aerospace vehicles and high-speed trains [8,9,24]. LM³ are architected structural systems that achieve lightweight, load-bearing, and multi-functional performance through deliberately designed topologies, hierarchical organizations, and spatial architectures [5]. Their distinctive characteristics arise mainly from topological configuration rather than chemical composition. Representative examples include structures with negative Poisson’s ratio, superelastic lattice systems, and elements with programmable deformation [5,6]. Mechanical metastructures have already been used in various engineering contexts, including impact protection, soft robotic actuators, spacecraft weight reduction, and biomedical implants [25-29]. In these applications, structural architecture becomes a primary design variable alongside material selection.

The design objectives of LM³ generally involve both performance enhancement and mitigation of conventional property trade-offs [3,4]. In practical terms, these structures are built to carry external loads while at the same time combining several functions in a single body: low weight, high load-bearing capacity and resistance to high temperature, creep, corrosion, fatigue, oxidation and erosion. On top of this basic requirement, further roles such as heat dissipation, impact protection, noise reduction, vibration isolation, and even morphing of the global shape can be added when needed [9,22]. Such multi-functionality originates from the ability of structural architectures to regulate tensile, shear, bending, torsional, and dynamic responses across multiple length scales. By tuning the generalized elastic response of the structure, one can obtain very large, or even formally negative, effective parameters. This design flexibility enables programmable mechanical behavior ranging from quasi-static deformation regulation to wave manipulation and vibration control [3-6].

LM³ represent a distinct class of architected lightweight structural systems, extending beyond the conventional concept of porous materials [20,21]. Beyond their superior static mechanical behavior, they offer significant advantages in dynamic performance for vibration, impact, noise and blast protection, as well as for thermal deformation control. These characteristics make LM³ particularly attractive for engineering applications involving impact loading, fatigue damage, dynamic fracture, and coupled multi-functional requirements, while maintaining high structural efficiency [25,26,29].

2 Mechanical Foundations and Design Framework for LM³

2.1 Mechanical benefits of lightweight multi-functional mechanical metastructures

A fundamental example in the study of mechanical metastructures is the comparison between stretch- and bending-dominated lattice structures. The Maxwell criterion and Gibson-Ashby models constitute classical frameworks for understanding the mechanical efficiency of cellular and architected structures. The Maxwell criterion distinguishes deformation mechanisms based on the relationship between nodal connectivity \mathit{\text{z}} and spatial dimensionality d (\mathit{\text{z}}⩾2d for stretch-dominated behavior), while Gibson-Ashby models describe the relationship between effective stiffness or strength and the relative density {\rho }^{\ast }/{\rho }_{\text{s}} power-law scaling [30,31]. However, additive manufacturing (AM) process-induced defects, such as incomplete nodal bonding and residual stresses, often lead to significant deviations from these theoretical assumptions. For example, selective laser melting (SLM)-fabricated metallic lattices frequently exhibit enhanced bending deformation due to reduced effective nodal connectivity {\mathit{\text{z}}}_{\text{eff}}, together with coupled tension-bending-shear responses, particularly in low-aspect-ratio struts with aspect ratios below 5 [32,33]. These effects reduce the predictive accuracy of classical models and indicate that idealized topology alone is insufficient to describe the actual mechanical response of manufactured metastructures.

Topology optimization provides a route beyond the simple Maxwell criterion by searching for lattice layouts with high and nearly isotropic stiffness, rather than relying only on predefined empirical patterns [34]. In parallel, programmable active lattices have been explored. A representative example is a lattice in which selected nodes respond to temperature, allowing the effective connectivity to change during actuation and enabling the structure to switch between stretch-dominated and bending-dominated modes [35]. Similarly, the classical Gibson-Ashby scaling laws, such as E^{\ast }/E_{\text{s}}\propto {\left({\rho }^{\ast }/{\rho }_{\text{s}}\right)}^1 for stretch-dominated behavior and E^{\ast }/E_{\text{s}}\propto {\left({\rho }^{\ast }/{\rho }_{\text{s}}\right)}^3 for bending-dominated behavior, often fail to accurately capture the stiffness of AM-produced metallic lattices because of multi-mechanism coupling, with experimental data typically lying between these two idealized bounds [31]. To address this issue, advanced multi-mechanism coupling models have been developed by incorporating weighting coefficients, such as {\alpha }_{\text{tensile}}, {\alpha }_{\text{bending}}, and {\alpha }_{\text{shear}}, into extended constitutive formulations:

(1)E^{\ast }=f\left({\rho }^{\ast },{\alpha }_{\text{tensile}},{\alpha }_{\text{bending}},{\alpha }_{\text{shear}}\right)

Further advances have been achieved through hierarchical structural designs, including bioinspired multi-scale lattices and locally heterogeneous reinforcement strategies. For example, titanium-alloy-reinforced nodes can partially decouple stiffness from density and have demonstrated yield strength improvements of up to 40% relative to conventional designs [36,37]. Multi-functional designs can also be realized through micro-perforated N-H-type sandwich structures [38,39]. By tailoring the pore layout, these structures can maintain relatively high load-bearing capacity while improving energy absorption and broadband sound absorption, with only a limited penalty in mass and overall volume [38,39]. These examples indicate that the mechanical benefits of LM³ do not arise from lightweighting alone, but from coordinated regulation of stiffness, deformation mode, energy dissipation, and functional response through architecture. Cross-scale modeling frameworks that couple molecular dynamics (MD) with continuum approaches, together with smart materials such as shape memory alloys and piezoelectric components, further provide possible routes for in-service mechanical reconfiguration and multiphysics co-design [40]. Representative mechanical benefits of LM³ are summarized in Table 1 according to their dominant design strategies.

2.2 Mechanical metastructure databases for lightweight composite structure optimization

Recent studies have shown that combining lattice-structure databases with topology optimization can improve sandwich-structure design and enhance multi-functional performance. For example, based on an equivalent-property optimization scheme, Jeong et al. [39] proposed a multi-morphology design route that led to an increase of about 86.9% in the flexural strength of composite sandwich panels. In their study, the improvement mainly came from coordinated changes in strut diameter from 0.5 to 2.0 mm and fillet radius from 0.1 to 0.5 mm, which produced controlled density gradients in the core. Following a similar idea, Wang et al. [40] developed a multi-scale design framework that explicitly considered the manufacturing constraints of laser additive manufacturing. Using this framework, they reported an approximately 40% increase in the stiffness-to-weight ratio of lattice-filled structures and showed, through tests on spacecraft load-bearing modules, that the design was experimentally feasible. Xiao et al. [42] further extended the gradient-core concept through concurrent optimization at both macro and micro scales, obtaining about a 220% improvement in the energy-absorption efficiency of hierarchical honeycomb cores. A related dynamic response model was then used to study impact loading and indicated a 35% improvement in impact-load distribution efficiency, even for geometrically asymmetric configurations [43].

A number of engineering studies already show that these methods are relevant to practical structural design. For example, topology optimization and experimental validation of aircraft wing ribs based on a Gulfstream G650 transonic jet showed that optimized rib designs could reduce mass by about 8%−15% relative to traditional wing-rib configurations [49]. Zhang et al. [44] produced 3D-printed AlSi10Mg lattice-core satellite panels and measured roughly a 2.1-fold increase in specific stiffness compared with conventional panel designs. From a process-informed design perspective, Jia et al. [50] set up a closed-loop selective laser melting (SLM) scheme in which the overhang angle was autonomously controlled in the range of 55° to 75°, reducing the amount of support material by about 70%. Other examples include the stress-line-mapped gradient lattices of Liu et al. [51], which showed an increase of about 280% in impact resistance, and the virtual-growth metastructures of Jia et al. [46], where the natural frequency was raised by about 22%, thereby alleviating some limitations of traditional periodic lattices. These results support database-assisted and topology-guided design as an effective route for next-generation sandwich cores, especially when mass efficiency, manufacturability, and multiple functional requirements must be considered simultaneously.

2.3 Major objectives of mechanical metastructures

Most of the progress in design, manufacturing, and database use outlined above is driven by three major objectives for mechanical metastructures. The first objective is to extend the accessible range of effective mechanical properties beyond conventional material-property bounds, for example, those given in the Milton-Ashby mechanics map [30,52]. The second objective is to reduce trade-offs among competing performance indicators, as typically summarized in the Gibson-Ashby materials map [26-29,31,52-55]. The third objective is to integrate multiple functions into a single lightweight structural system [20,23,25]. The intrinsic mechanical characteristics of natural materials and artificially fabricated alloys or composites are typically described by their elastic constants and are commonly visualized using two principal diagrams: the Gibson-Ashby plot, which relates Young’s modulus to density [31], and the Milton-Ashby plot, which illustrates relationships between bulk and shear moduli [54]. A central goal of LM³ is therefore to create effective property combinations that approach or exceed the conventional regions represented in these canonical maps [17,39,45].

Conflicts among multiple mechanical property indicators are common, as exemplified by the well-known strength-toughness trade-off [26,54]. Overcoming such contradictions requires the customized design of structural genomes, unit-cell architectures, and multi-scale topological connectivity [19,42,46]. The mechanical role of each structural element, such as a strut or node, can be encoded in the corresponding generalized elasticity matrix and adjusted according to functional requirements [1,17,26-29,52-55]. By precisely configuring the connections among neighboring components, different mechanical responses can be selectively decoupled [25]. This strategy enables the extension of performance limits, the realization of mechanical behaviors absent in natural systems, and the coordinated activation of multiple deformation mechanisms for synergistic multi-objective optimization of macroscopic performance [27,28]. In this way, mechanical metastructures provide a structural route for mitigating long-standing conflicts among stiffness, strength, toughness, recoverability, energy absorption, and multi-functional performance.

As summarized in Table 2, the three objectives of mechanical metastructures are closely related rather than independent. Extending the accessible property range usually requires architectural mechanisms that are difficult to obtain in homogeneous materials, such as negative stiffness, recoverable buckling, hierarchical damping, or recoverable high-strength nanolattices [29,56-58]. Mitigating performance trade-offs relies on a different but related mechanism: deformation, damage, and energy dissipation are redistributed across struts, nodes, interfaces, pores, or multiple length scales, so that one property can be improved without a proportional loss in another [58-61,64]. Multi-functional integration further extends this idea by assigning different mechanical, thermal, acoustic, transport, or electrochemical roles to different architectural features within the same load-bearing structure [65-69]. This perspective shifts the design problem from selecting a material with fixed intrinsic properties to designing an architecture that controls how forces, deformation, and energy are transmitted through the structure. For example, stiffness and recoverability can be balanced through negative-stiffness or recoverable lattice mechanisms [29,56,57]; strength and toughness can be improved through sacrificial interfaces or damage-delaying coatings [58,59]; and lightweight structures can achieve enhanced energy absorption by using hybrid TPMS, hierarchical, or aperiodic topologies [60,61]. Similarly, surface robustness, thermal management, acoustic absorption, and structural energy storage can be incorporated when the architecture is designed to separate or coordinate multiple functions at different length scales [65-69].

Therefore, the mechanical benefits of LM³ are not limited to achieving a single high property index. Their broader value lies in creating property combinations that are difficult to realize in conventional materials or monolithic structures, including high stiffness with damping, strength with recoverability, lightweighting with energy absorption, and load-bearing capacity with thermal, acoustic, or electrochemical functions [29,56-69]. The following section further organizes these ideas through a tetrahedral framework, before the review turns to structural genomes, data-driven models, and AI-assisted design methods.

2.4 Tetrahedral paradigm of mechanical metastructures

Current research on mechanical metastructures can be broadly organized around four interrelated aspects. The first direction concerns the physical mechanisms and governing principles of mechanical metastructures. It focuses on how topology, connectivity, hierarchy, and deformation modes can be used to achieve unusual mechanical properties or to mitigate conflicts among multiple physical attributes [3-6]. The second direction focuses on structural genome libraries at the level of local structural primitives, unit cells, and their combinations. Representative sources include alloy-inspired architectures [29], bioinspired designs [55-58], topology-optimized configurations [42,43,45,46,50], and AI-generated geometries [1,2,19]. The third direction concerns metastructure design theory and performance optimization, including homogenization methods, topology optimization, structure-property models, multi-scale integrity analysis, and design approaches for extreme environments and coupled multiphysics conditions [4,22,42,43]. The fourth direction involves fabrication, validation, and service assessment, including scalable manufacturing of metallic, polymeric, ceramic, and composite metastructures, characterization of manufacturing defects, process repeatability, high-fidelity testing, and long-term performance evaluation under realistic service conditions [21,25-27,32,33,40].

By analogy with the classical materials tetrahedron that captures the composition–structure–process–property relationship, the proposed tetrahedral framework describes mechanical metastructures through four coupled dimensions: structural genome, design theory, function/property optimization, and fabrication/validation. The structural genome dimension defines the basic architectural variables of metastructures, including unit-cell topology, nodal connectivity, strut or shell geometry, chirality, hierarchy, gradients, and spatial arrangement. These variables serve as the structural basis for regulating stiffness, strength, deformation mode, energy absorption, and wave propagation [30,31]. The design-theory dimension links these architectural variables to measurable responses, such as effective modulus, Poisson’s ratio, buckling resistance, impact resistance, fatigue response, dynamic stability, and thermal deformation. This connection is established through homogenization theory, topology optimization, multi-scale modeling, interface design, and structure–property relationships for architected systems [4,42,43].

The function and property dimension emphasizes that mechanical metastructures are not designed only to maximize a single mechanical index. Instead, they aim to balance lightweighting, load-bearing capacity, heat dissipation, energy absorption, sound absorption, vibration attenuation, thermal expansion control, and other service-oriented functions within the same structural system [20-23,25-27]. The fabrication and validation dimension determines whether the designed architecture can be reliably manufactured and whether its measured performance is consistent with theoretical or computational predictions. For this reason, manufacturing routes, defect tolerance, dimensional accuracy, process repeatability, and experimental verification are not secondary issues, but essential components of the metastructure design framework [32,33,59].

As shown in Fig. 1, the integration of artificial intelligence with mechanical metastructures drives an evolution from the classical tetrahedron to an intelligent twin-tetrahedron paradigm [27-29]. In this extended framework, structural genomes, design theory, fabrication processes, and functions/properties are no longer connected only through static design rules. They are increasingly linked through data-driven models, physics-informed learning, high-throughput simulations, automated experiments, and service feedback. Data practices consistent with findable, accessible, interoperable, and reusable (FAIR) principles support the organization and reuse of metastructure data, while AI-enabled methods assist structural genome generation, performance prediction, process optimization, and closed-loop correction [27-29,46,50]. Compared with the classical tetrahedron, the twin-tetrahedral framework places greater emphasis on measurable variables, data traceability, manufacturing feedback, and performance verification. This provides a clearer basis for the subsequent discussion of AI-assisted design, fabrication optimization, and service-oriented deployment of mechanical metastructures.

3 Fusion of AI and Mechanical Metastructures

3.1 Motivation and classification of AI-enabled mechanical metastructure design

The use of artificial intelligence in scientific research can be understood as an extension of earlier research paradigms, including empirical observation, theoretical modeling, computational simulation, and data-driven discovery [1,70,71]. Empirical studies accumulate knowledge through observation and experimentation, while theoretical models describe governing mechanisms using mathematical and physical formulations. Computational simulation further extends this capability by allowing complex problems that are difficult to solve analytically to be investigated at larger scales and with higher fidelity. Data-driven methods then use experimental, computational, and literature-derived datasets to identify correlations, construct surrogate models, and guide further modeling or experimentation. Within the current AI for Science paradigm, machine learning and related AI tools are increasingly used to support hypothesis generation, model acceleration, design optimization, and knowledge discovery.

Mechanical metastructures still face a number of fundamental difficulties, many of which arise from strong nonlinear effects in design, fabrication, and functional control [5,6,30]. At the design level, the overall response is determined not only by the properties of the constituent materials, but also by the interactions within the microarchitecture. These interactions often produce nonlinear constitutive behavior, making both forward prediction and inverse design challenging [3,4,72]. On the manufacturing side, maintaining geometric accuracy and process repeatability remains difficult, especially for three-dimensional architectures with fine features. Small geometric deviations or material inhomogeneities at the micro scale can lead to noticeable differences between measured performance and the values predicted by simplified theoretical models [31-33]. For functional regulation, nonlinear mechanisms such as buckling or phase transitions are attractive because they can generate multi-functional responses, but reliable methods for controlling critical states and programming dynamic responses remain limited [53,57,73]. These issues require cross-scale nonlinear models, improved manufacturing control and characterization, and data-assisted strategies that can be validated against physical experiments [1,19,40].

Multi-scale design and simulation of architected mechanical metastructures are often governed by systems of differential equations across multiple length scales. Two difficulties are particularly important. First, constitutive relations are often uncertain or incomplete: vibration, deformation, heat or mass transport, and fluid flow may involve different governing equations and closure assumptions. Second, these equations are usually nonlinear and high-dimensional, which makes direct numerical simulation computationally expensive and sometimes unstable [15,72,74]. Compared with purely physics-driven design strategies, AI-based methods can provide efficient surrogate models, identify structure–property correlations, and support inverse design in high-dimensional design spaces [1,2,75]. However, AI should not be viewed simply as a replacement for mechanics-based modeling. For mechanical metastructures, its value lies mainly in combining data, physical constraints, and design objectives to improve prediction, optimization, and validation [15,16,76,77].

Although large-scale data-driven material models have attracted increasing attention, many of them remain predominantly empirical and rely mainly on statistical fitting rather than physically constrained or physics-informed formulations [16,76,77]. As summarized in Fig. 2, AI-enabled multi-scale simulation and mechanical metastructure design can be classified according to the role of data and physics in the modeling workflow. At one end are physics-based multi-scale models, which rely mainly on established theories and require limited training data. A second category uses machine learning to accelerate simulation or data assimilation. A third category, represented by physics-informed neural networks, embeds governing equations or physical constraints into the learning process [15,16,77]. A fourth category consists of data-driven deep learning models combined with feature engineering, which are particularly useful when large structural or image-based datasets are available [78-80].

Against this background, machine learning, deep learning, and related techniques have been increasingly used in mechanical metastructure design, property prediction, fabrication optimization, and service-performance evaluation [75,81,82]. For example, neural networks can be trained to infer unit-cell architectures from prescribed target properties, allowing generative design methods to explore candidate structures beyond conventional manual design routes [82-84]. To clarify the methodological landscape, Table 3 summarizes representative AI methods according to their modeling strategy, typical applications, and applicability limits. Table 4 further groups representative software platforms and design tools according to their main functions, with detailed source information provided through the corresponding references. These tables are intended to provide a compact reference for the following subsections, rather than to replace critical discussion. Specifically, the following sections examine data-driven high-throughput design, physics-informed machine learning, AI-assisted integrated computational metastructure engineering, inverse and generative design, and AI-assisted bioinspired metastructures.

3.2 Data-driven high-throughput intelligent design of mechanical metastructures

The data-driven design paradigm for mechanical metastructures establishes a systematic workflow that spans data collection, feature extraction, model construction, and optimization-driven application [1,19,75]. The process typically begins with the compilation of structural and performance datasets obtained from experiments, numerical simulations, or hybrid experimental-computational pipelines [80,122,149]. Once a suitable set of descriptors has been chosen, these features can be used to build machine-learning or deep-learning models that link structural architecture to mechanical response [74,76,77,156]. After training and validation, such models can serve as fast surrogates, allowing candidate structures to be screened or optimized without repeatedly performing expensive simulations or experiments [2,78,84,108,112,124]. In this sense, data-driven design is useful not simply because it accelerates computation, but because it provides a practical route for exploring high-dimensional relationships among topology, geometry, material, process, and performance [75,140,151].

Recent work has improved this workflow mainly through larger datasets, graph-based representations, image-based learning, and latent-space modeling. Large simulated datasets, in some cases approaching the billion-sample scale across multiple material systems, have enabled the training of deeper networks with improved generalization [104]. Graph neural networks have been applied to shell-type metastructures to capture geometry–property relationships [148], while data-driven models have been used to predict complex and partially chaotic behavior in multistable origami systems under noisy conditions [162]. Computational efficiency has also improved: certain deep-learning schemes can estimate effective isotropic elastic properties directly from structural images with about 90% accuracy and millisecond-scale inference times [132]. In addition, variational autoencoders (VAEs) combined with property regressors can map complex microstructures into a low-dimensional latent space, where effective properties can be adjusted by manipulating latent variables [163]. Digital structured-genomics approaches that combine machine learning with finite-element calculations have also been proposed for designing anisotropy and spatial distributions of material within a given volume [134].

Data-driven tools have also changed the way structural optimization is performed. Machine-learning models combined with genetic algorithms can accelerate the search for improved layouts in complex systems such as auxetic honeycombs [77]. Bayesian regression has been introduced into multi-objective optimization and used, for example, to design cylindrical lattice shells with approximately twice the buckling capacity of conventional designs [166]. Related approaches have supported aperiodic tree-like support structures [164], two-dimensional lattice cells with locally adjusted properties [165], and functionally graded materials obtained through self-learning design procedures [167]. High-throughput data-driven design can therefore broaden the searchable architecture space of mechanical metastructures. At the same time, its reliability depends on descriptor quality, data coverage, physical consistency, and validation against manufacturable structures; otherwise, the resulting models may be computationally efficient but difficult to generalize.

3.3 Physics-informed machine learning in mechanical metastructure design

The use of physics-informed neural networks (PINNs) in mechanical metastructure design usually follows a physics-constrained learning workflow. One first sets up a physical description of the problem, in which initial and boundary conditions and basic conservation laws, such as momentum or energy balance, are written in a form that can be imposed during model training [15,16,77]. A neural network is then constructed under these constraints and, where possible, trained with additional experimental or simulation data [16,76,77]. During training, the model is adjusted so that its outputs are consistent both with the governing equations and with the available observations. Once this stage is completed, the network can be used as a fast surrogate for performance prediction and for exploring alternative layouts, provided that the governing equations, boundary conditions, and training data are sufficiently representative of the target design problem [15,16]. In this sense, physics-informed learning is not simply a data-fitting strategy; it attempts to embed mechanical consistency into model training and prediction, which is particularly important for nonlinear, multiscale, and defect-sensitive metastructures [15,16,72].

A natural extension is to combine physics-informed learning with Bayesian machine learning. Their built-in treatment of noise and uncertainty makes it easier to quantify confidence in the predicted response, which is important when the available data are limited, noisy, or heterogeneous [86,94]. In one example, this approach has been used to design defect-sensitive metastructures. At the macroscopic level, systems have been reported that reach about 94% compressive strain with a recoverable strength of roughly 0.1 kPa. In contrast, at the microscopic level, recoverable strengths above 100 kPa have been obtained together with about 80% compressive strain [86]. Such examples illustrate the value of combining physical constraints with uncertainty quantification, especially when the measured response is sensitive to defects, scale effects, or limited training data [86,94,98].

The combination of physical constraints with data-driven models has also benefited from specialized network architectures and image-based data. One example is the use of U-Net-type convolutional networks built on computed tomography (CT)-based three-dimensional reconstructions. In this case, the network is trained to identify and remove powder particles attached to the surface of additively manufactured aluminum parts. The cleaned geometries are then passed through a surface-smoothing step so that they can be meshed more reliably for finite-element analysis and used in subsequent compression simulations [146]. In a related direction, convolutional neural network (CNN) architectures have been proposed to design customized lattice metastructures by learning from datasets of pseudorandom pillar geometries and their corresponding stress-strain responses, demonstrating strong generalization beyond the original training set [136]. These studies show that physics-informed machine learning can operate not only on scalar descriptors, but also on geometric, voxel, image, and field data generated by experiments or simulations [136,146,149,156].

Machine learning has also been applied to additive manufacturing, where classification, regression, and clustering have been used to create high-performance metastructure designs, optimize process parameters, monitor defects, and support manufacturing planning and quality control [186]. For instance, CNN-based optimization schemes have been developed to map local sound fields to metasurface phase gradients, surpassing genetic algorithms in accuracy while achieving simultaneous determination of phase gradients and regional sound-field control [140]. Deep learning approaches have likewise been introduced to obtain accurate full-field predictions of strain and stress tensors in composite geometries, with model architectures that are explicitly constructed to respect the fundamental laws of continuum mechanics [155]. These examples indicate that physically constrained learning is particularly useful when the target output is a field quantity, such as strain fields, stress fields, acoustic phase distributions, or defect maps, rather than a single scalar property [140,155,156].

Further examples show how imaging and data-driven tools can be combined with mechanics-based models. Scanning electron microscopy (SEM) and CT have been used to characterize geometric defects in nickel-based microlattice structures; by feeding this information into random-field descriptions together with Monte Carlo simulations and finite-element analysis, critical buckling loads and strength scatter can be estimated with reasonable accuracy [187]. Convolutional neural networks have also been trained for property optimization using data generated by energy homogenization schemes [139], and hybrid approaches that link genetic algorithms with CNNs have been proposed for multi-objective microstructure design under complex loading states [139]. Physics-informed machine learning is most effective when mechanical constraints, geometric data, uncertainty estimates, and validation experiments are considered together [137,139,146,187]. Its limitations should also be recognized: model performance can be sensitive to incomplete physical descriptions, inconsistent boundary conditions, biased training data, and insufficient validation for manufacturable three-dimensional architectures [15,16,76,77,186].

3.4 AI-assisted integrated computational metastructure engineering

AI-assisted integrated computational metastructure engineering can be viewed as an extension of integrated computational engineering to architected mechanical systems, in which artificial intelligence is used to connect structural genomes, multi-scale simulations, experimental measurements, and macroscopic performance evaluation [163,167]. In practical terms, machine learning, deep learning, multi-scale computational simulations, and experimental measurements are combined so that behavior can be predicted and optimized across several scales, from structural genomes and unit cells to cross-scale metastructures and their macroscopic properties [156,161]. A typical workflow begins with multi-source data collected from experiments, numerical simulations, or reconstructed structural geometries [149,187]. AI algorithms are then used to identify structure–property correlations in these datasets [134,141], which are in turn used to build models for macroscopic performance [129,156]. These models can be coupled with physical principles and multi-scale calculations [124,166], and finally applied to accelerate the design and optimization of mechanical metastructures [108,140]. For example, generative adversarial networks (GANs) have been used to design complex architectures with tailored mechanical properties [188]. In this setting, extensive datasets of structural architectures are generated, adversarial networks are trained to synthesize and classify candidate architectures, and new material designs that meet prescribed target properties are proposed. Similarly, GAN-based algorithms have been employed to design topological configurations of sound-absorbing porous materials. Training datasets constructed via finite-element simulations enabled design processes that were accelerated by several hundred times relative to conventional methods while preserving broadband sound absorption performance [157]. These examples show that AI-assisted computational frameworks are particularly useful when the design space is too large for direct enumeration, but the generated structures still require physical constraints, manufacturability checks, and experimental or numerical validation [157,188].

As illustrated in Fig. 3, and in analogy with materials informatics, the upper route shows that the properties of mechanical metastructures can be predicted using data-driven paradigms supported by structural descriptors, feature engineering, and machine-learning algorithm libraries [110,114]. Such approaches are useful when sufficiently representative structural descriptors or image-based data are available, because they can provide rapid property estimates without carrying out high-fidelity multi-scale simulations for every candidate design. In contrast, the lower route shows an AI-assisted multi-scale prediction and inverse-design workflow, in which multimodal datasets and mechanics knowledge are combined with AI models and integrated computational structural engineering simulations [149,166]. This route starts from established composite theories and domain knowledge, with the required datasets added where needed. Machine-learning tools are then brought in at selected structural scales, mainly to speed up simulations, improve model updating, and support inverse design rather than to replace the underlying mechanics models [124,142].

These two routes are complementary. Data-driven schemes are efficient for rapid screening, inverse search, and property prediction, whereas AI-accelerated multi-scale simulations retain stronger links to mechanics-based modeling and are better suited for problems involving limited data, strict physical constraints, or complex service conditions. For mechanical metastructures, the most reliable design strategy is therefore not a purely data-driven replacement of mechanics, but an integrated workflow in which data, physical models, simulations, and validation experiments are used together to improve prediction accuracy, design efficiency, and engineering reliability [124,142,149,166].

3.5 Inverse design, generative design, and multi-objective optimization

Inverse design of mechanical metastructures is mainly used when the target mechanical response is clearly defined and the performance requirements are stringent [111,121,124]. Instead of starting from a given layout, one first specifies the desired response and then searches for an architecture that can realize it. The structure-property relationship can be described by either a physics-based model or a data-driven surrogate, and optimization tools, such as genetic algorithms or gradient-based methods, are then used to identify configurations that meet the prescribed target properties [138,151]. The effectiveness of inverse design therefore depends on three factors: the accuracy of the forward model, the robustness of the optimization algorithm, and the physical validity of the structure-property relationship used to guide the search.

Generative design follows a different route. It typically relies on datasets that contain structural configurations and their associated properties. Generative models, including GANs and VAEs, are used to learn the statistical distributions of structural genomes, unit-cell features, and macroscopic physical properties from these datasets [140,157]. High-throughput screening can then be used to identify promising candidates, which may be further refined using targeted optimization. Compared with conventional inverse design, generative design is better suited for exploring large and non-intuitive design spaces, but it also places stronger demands on data quality, descriptor consistency, manufacturability constraints, and model generalization.

The combination of inverse and generative design strategies has led to several advances in mechanical metastructures. For example, an inverse machine-learning framework based on GANs has been developed to generate optimized lightweight lattice structures with 40%–120% higher load-bearing capacity than traditional octet-truss architectures [2]. Neural networks incorporating physical constraints have also been used to design spatial topologies of structural components and lattice architectures with customized anisotropic stiffness distributions [97]. In addition, the integration of deep learning with heuristic optimization algorithms has enabled multi-functional magneto-active metastructures with on-demand control of deformation, stiffness, acoustic band gaps, and electromagnetic responses [186].

These methods have also been extended to complex mechanical behaviors and application-specific structures. Artificial neural networks trained on finite-element simulation datasets have enabled the inverse design of strain-hardening, strain-softening, and stable-crushing responses, together with controlled structural collapse processes [99]. Three-dimensional convolutional neural networks have been used to rapidly predict the mechanical properties of grid architectures, supporting the inverse design of systems with enhanced anisotropy and improved connectivity. Compared with conventional methods, these approaches achieved approximately 10% improvements in design accuracy and stiffness [118]. Related methods have been further applied to mechanical metastructures for energy focusing, dispersion control, dynamic wave manipulation, and nonlinear motion conversion [123].

Applications of inverse and generative design now extend across several engineering contexts. Generative computational frameworks have been used to produce irregular mechanical metastructures that regulate stress distributions within structural components and unit cells, including uniform global stress distributions, localized stress amplification, and targeted reversal of local stress directions [46,112]. In footwear applications, inverse design has been used to develop structured midsoles with spatially varying dynamic responses, so that different regions follow prescribed stress-strain curves [100]. Large experimental datasets combined with deep neural networks and genetic algorithms have also been used to guide the design of buckling lattice structures, where improved buckling strength has been reported for rib-type and bioinspired grid layouts relative to conventional periodic lattices [103].

Recent developments further indicate that inverse and generative design are moving from scalar-property optimization toward spectrum-, field-, and function-oriented design. For example, inverse design methods have been used for phononic metastructures, where band gaps can be tuned more systematically and efficiently [189]. Transformer-type networks, such as AcoustoGPT, have also been applied to the inverse design of acoustic metastructures with prescribed sound-absorption spectra [190]. Future inverse and generative design frameworks should therefore combine architecture search with physical consistency, manufacturability, robustness, and multi-objective performance under realistic service conditions.

3.6 Integration of AI and bioinspired mechanical metastructures

The combination of artificial intelligence with bioinspired structural design provides a systematic route for translating biological architectures into engineering metastructures. Instead of relying only on empirical trial-and-error, design can be guided by algorithms that extract, simplify, and recombine structural features observed in natural systems. By integrating computational modeling, advanced fabrication methods, and deep-learning tools, bioinspired design can be extended from qualitative imitation toward data-assisted structure generation and performance optimization across multiple length scales [1,2,55-58].

Deep generative models, such as GANs and VAEs, are useful for representing biological architectures that are difficult to describe using simple geometric rules. Representative examples include the crossed-lamellar structure of Strombus gigas shells, where four types of lamellar elements form a three-dimensional architecture with high fracture resistance [191], and porous cuttlebone, whose asymmetric S-shaped walls and lamellar septa provide high energy-absorption capacity [192]. These biological systems have inspired mechanical metastructures that aim to mitigate strength–ductility and strength–density trade-offs through hierarchical layouts. Mechanically guided assembly has been used to transform two-dimensional films into three-dimensional curved mesosurfaces with controlled porosity and curvature [193], while AI-based design tools have generated nonperiodic three-dimensional architectures with predictable direction-dependent elastic behavior [194]. Related advances have also extended architected mechanics into biomedical systems, where geometry can regulate cell behaviors, tissue responses, and device performance [195].

Bioinspired mechanical metastructures are also relevant to dynamic response control and adaptive functions. In this context, biological adaptive mechanisms provide design cues for synthetic systems that combine responsive materials, programmable geometry, and learning-based control. For example, crumple-recoverable electronics inspired by butterfly wing emergence can reversibly switch between a soft state of about 2 MPa and a rigid state of about 1.3 GPa while maintaining nearly complete shape recovery [196]. Related concepts have been explored in biomedical devices, such as cardiac patches with a negative Poisson’s ratio that accommodate myocardial strains of about 40% [197]. The hierarchical hinge structure of bivalves, consisting of radially aligned aragonite nanowires embedded in a softer matrix, provides another example of fatigue-resistant architecture, with reported stability over hundreds of thousands of loading cycles [198]. Bioinspired three-dimensional flexible devices further broaden the design space for systems requiring miniaturization, heterogeneous integration, and coupled mechanical and electronic functions [199].

A further direction is multiphysics integration, where bioinspired architecture is combined with topology optimization and additive manufacturing to coordinate mechanical, mass-transport, chemical, and biological functions. For example, a wood-inspired catalytic metastructure with overlapping microlattices and a bimodal pore network can partially decouple stiffness, fluid transport, and catalytic activity, increasing mass-transport efficiency by up to about 400% without an apparent loss of structural integrity [55]. Enamel-like composites with hierarchical nanowire–microbundle–macroarray structures have achieved fracture toughness values about 3.4 times those of natural enamel [200], and AI-designed heterogeneous architectures have been proposed for tunable anisotropic stiffness profiles [201]. Similar principles appear in Mg–Ti interpenetrating-phase composites with brick-and-mortar, Bouligand, or crossed-lamellar architectures, which favor stress transfer, damage delocalization, and crack arrest [202]. These studies indicate that AI-assisted bioinspired design is most valuable when it moves beyond reproducing biological forms and instead identifies transferable architectural principles. Engineering deployment will require manufacturable architectures, defect-tolerant designs, scalable fabrication routes, and validation under realistic service conditions.

4 Development Trends and Future Directions of Mechanical Metastructures

4.1 Digital twin design for tailoring multi-scale topology of mechanical metastructures

Digital twinning provides a natural framework for connecting the design, fabrication, testing, and service assessment of mechanical metastructures. For LM³, this connection is particularly important because the final performance depends not only on the designed topology, but also on cross-scale structural features, manufacturing deviations, and service-induced changes. Therefore, digital twin design should be understood not as a post-processing visualization tool, but as a closed-loop method for linking virtual structural models with experimental and service data.

The multi-scale design of LM³ is an iterative, bottom-up process that relies on close cross-scale coordination among material properties, unit-cell architectures, mesoscale topology, and macroscopic structural performance. Its central objective is to establish precise links between microstructural features and macroscopic performance through cross-scale modeling and performance transfer [72]. The process typically begins at the atomic or molecular scale, where first-principles calculations or MD simulations are used to probe the intrinsic mechanical properties of the base materials. At the micro- and mesoscopic scales, computational homogenization methods or machine-learning surrogate models are employed to parameterize designed topologies, such as truss or chiral architectures, and to obtain equivalent macroscopic elastic tensors, Poisson’s ratios, and other effective properties [74,124]. At the macroscopic scale, these equivalent properties are assigned to periodically or spatially arranged metastructures, and finite-element analyses are carried out to evaluate their overall mechanical responses. Key enabling technologies include topology optimization for automated generation of microstructures [42,46], machine learning for accelerating cross-scale correlations and inverse design [80,151], and additive manufacturing for realizing virtual multi-scale designs and validating them experimentally [197,200,203,204].

For major industrial equipment, LM³ must usually satisfy several requirements at the same time, including low weight, load-bearing capacity, environmental resistance, fatigue durability, impact resistance, vibration isolation, and thermal deformation control [25,54,80,205-212]. Within a digital twin framework, these requirements are first translated into structural-genome composition, unit-cell design targets, and performance constraints. Candidate metastructures are then screened, modified, and optimized using structural-genome libraries, data-driven models, generative design methods, and quantitative structure–property relationships [39,80,100,150,202,208,213-215]. The final stage is customized mechanical design for specific application scenarios, where the virtual model, manufacturing constraints, and measured performance are iteratively compared and updated [25,36,100,205,216,217]. This closed-loop workflow is more suitable than a purely bottom-up design route because it allows multi-scale topology, processing constraints, manufacturing deviations, and service requirements to be considered within the same design cycle.

Moreover, it is essential to evaluate the structural safety and durability of LM³ using digital twin technologies together with service-performance validation experiments on representative structural components [149]. A digital twin for mechanical metastructures integrates physical entities, sensor data, numerical simulations, and AI-based model updating. Its key technical characteristics include high-fidelity virtual mapping of structural features to performance [74], real-time bidirectional data interaction through embedded sensing and model calibration [148], and closed-loop performance prediction and optimization in virtual service environments [80]. The main advantage is that the conventional design–build–test route can be transformed into a predictive and continuously updated workflow: the number of costly experiments can be reduced, design cycles can be shortened, and in-service behavior can be monitored and forecast over the component lifetime [160].

This framework also connects naturally with embodied intelligence, where the structure itself contributes to sensing, actuation, and adaptation [62]. Possible applications include bioinspired robots, wearable exoskeletons with adjustable stiffness, and built-in sensing elements that provide self-feedback on mechanical states. In these systems, the mechanical response is programmed through microarchitecture, while shape and stiffness can be adjusted by thermal, magnetic, or electrical stimuli [196]. Embedded sensing can further couple local deformation with electrical or optical signals, providing self-sensing capability [25]. Porous and architected topologies help maintain low weight and low energy consumption while preserving the required mechanical function [28]. Thus, digital twin design and embodied intelligence together point toward LM³ systems that can be designed, validated, monitored, and adapted under realistic service conditions.

4.2 Structural Genome Initiative framework for high-throughput metastructure design

For large-scale use, metastructures require fabrication methods that can produce periodic, hierarchical, or sub-wavelength features over wide areas at reasonable cost and speed. In practice, nanoimprint lithography (NIL) and self-assembly (SA) are two of the main routes that have been explored for optical metastructures and metasurfaces [218,219]. NIL offers high-resolution patterning for active metasurfaces and can be scaled up with suitable tooling [218], while the bottom-up character of SA allows more complex nanostructures to form, some of which are difficult or impossible to obtain by conventional top-down approaches [219]. On this basis, several more advanced platforms have been developed to extend the range of achievable structures. Multi-focus two-photon lithography, for example, combines femtosecond laser amplifiers, specially formulated photoresists, and holography-based digital micromirror device (DMD) scanners to increase throughput in the fabrication of three-dimensional metastructures with intricate topology [220]. Surface-tension-driven self-assembly has been used in a different way, making use of lipidic cubic phase morphologies and simple energy-minimization ideas to generate new types of metastructure [221]. In addition, origami-inspired folding has been coupled with automated roll-to-roll processing to turn flat magnetic sheets into three-dimensional soft magneto-active devices with tailored geometries and multiple deformation modes [222].

In parallel with fabrication advances, bioinspired design has also been used to ease some of the usual performance trade-offs. Using Douglas fir wood as a design template, metastructure catalysts with microlattice cores have been proposed that balance mechanical strength, mass transport, and catalytic activity more effectively; in these systems, fluid velocity and reaction efficiency can be increased without a severe loss of stiffness [55]. High-throughput experimentation combined with computation has further accelerated the search for suitable architectures. Using large experimental datasets, iterative loops between simulation and test now make it possible to pick out metastructures with tailored mechanical responses in a more systematic way [223].

These advances also provide a foundation for so-called Structural Genome Initiative (SGI) frameworks, which aim to link material composition, structure, and properties more systematically. For example, SGI-based deep neural networks have been trained to predict basic mechanical properties of three-dimensional woven lattices from a small set of genomic descriptors [224]. SGI simulation frameworks go a step further by combining materials-genome concepts with structural analysis, leading to constitutive models that can describe multi-scale composites with relatively high fidelity [225,226]. Overall, these SGI-type approaches suggest a more organized route for the design and manufacture of advanced metastructures, bringing fabrication, characterization, and computation under a single, coherent design picture.

As summarized in Fig. 4, the SGI-based high-throughput research system for LM³ comprises four main components: a multimodal database for mechanical metastructures [227]; high-throughput intelligent computing and manufacturing process optimization [228,229]; control software for high-throughput intelligent automated experimental platforms [230,231]; and high-throughput intelligent automated experimentation [230,232,233]. Beyond these four components, the SGI framework also requires interfaces for automated device control, multimodal process monitoring, intelligent data analysis, and process-parameter optimization [234-237].

In the construction of the database and intelligent data-analysis pipeline, high-throughput preparation and formulation data for structural genome primitives and unit cells are collected [227], together with multimodal in situ monitoring data from the manufacturing process [235]. Real-time acquisition covers manufacturing process variables, defect-characterization data, spatial-topology data, and macroscopic performance testing results [230,232]. On the basis of these multimodal and multi-scale datasets, intelligent segmentation, numerical reconstruction, and high-fidelity modeling of manufacturing defects and structural topology are carried out using nondestructive testing technologies [238]. Artificial intelligence algorithms, including multimodal data fusion, graph convolutional networks [238], and transfer learning [236], are then used to construct multi-scale models that correlate spatial topological features with macroscopic performance. In situ experiments and characterization procedures reveal the evolution of topological features and the influence of defects on performance [235], thereby guiding high-throughput manufacturing optimization [230].

Within the intelligent control software, new principles and methods for high-throughput metastructure manufacturing are developed [239]. These include gradient structural genomes, gradient unit-cell geometric parameters, and orthogonal experimental strategies for screening combinations of process parameters. Key process factors are identified, and control directions and incremental steps are optimized for primitives, unit-cell gradients, and multi-scale spatial topologies [228,229], enabling stable control and reliable fabrication of topological and geometric gradients. In the high-throughput preparation device, strategies for stable evolution control of structural primitives and spatial topology are embedded into the intelligent automation control software [231], allowing real-time adjustment of experimental parameters [230]. Recent control software has increasingly incorporated near-real-time monitoring, automated feature extraction, and adaptive process adjustment based on multimodal process signals and metastructure characterization data [232,235]. This allows automatic extraction of key process features, defect characterization over different length scales [238] and, in some cases, on-the-fly adjustment of process parameters during high-throughput runs. The resulting closed loop supports adaptive adjustment of manufacturing settings and more targeted design of subsequent experimental cycles [231,239]. In practice, this improves the overall efficiency of preparation, characterization, and testing [228], while at the same time building up large datasets on both process parameters [233] and performance [227].

When combined with machine learning models [236], the system enables intelligent prediction of microstructural attributes, defect characteristics, and their evolution from process parameters [238], supporting the establishment of predictive models that link structural primitives and unit cells, integrated manufacturing processes, spatial topology, and macroscopic performance [240]. Ultimately, through the accumulation of high-throughput experimental data, fusion of experimental and simulation datasets [227] and machine-learning-based mapping of process-topology evolution [236], optimized fabrication processes for mechanical metastructures can be derived [229]. Together, these elements form an intelligent control framework that links structural genomes, unit-cell design, manufacturing processes, spatial topology characterization, and performance evaluation. Within the SGI framework, this enables autonomous perception–analysis–execution cycles for high-throughput fabrication and characterization [230,231].

4.3 Large language model-based multi-agent systems for closed-loop metastructure design

Large language model (LLM)-based multi-agent systems are relevant to metastructure research because the design process usually involves several coupled tasks, including literature analysis, data processing, simulation execution, design comparison, and experimental planning. Compared with training small task-specific models for LM³ from scratch, domain models distilled from larger foundation models may improve accuracy and efficiency when sufficient domain data and validation procedures are available [241,242]. In a multi-agent workflow, different agents can be assigned to literature retrieval, dataset cleaning, simulation execution, design generation, and experimental planning. Such a division of tasks allows LLMs to be used as interfaces between domain knowledge, computational tools, and experimental data, rather than as independent design engines.

Recent general-purpose LLMs that may serve as backbone models for these workflows include proprietary models such as GPT-5.5, GPT-5, Claude Fable 5, Claude Opus 4.8, Gemini 3.5, and DeepSeek-V4, as well as open or open-weight model families such as Qwen3, Qwen3-Coder, Llama 4, Mistral Large 3, Mistral Medium 3.5, and Mistral Small 4 [243-249]. These models differ in reasoning tasks, multimodal processing, code generation, tool use, context length, deployment cost, data-governance requirements, and compatibility with local or cloud-based research infrastructure. For LM³, model selection should therefore be based on structure interpretation, reliable source retrieval, controllable tool invocation, scientific computing capability, and integration with simulation or experimental platforms, rather than on general benchmark scores alone [250,251]. Recent studies have begun to connect LLMs with generative models and domain-specific simulation tools for materials and metastructure design. These frameworks are not tied to a fixed model combination; they can be implemented using different foundation models, geometry generators, code agents, and simulation tools according to the target task. CrossMatAgent, for example, has been proposed as a hierarchical multi-agent framework for metastructure design, in which multimodal reasoning, structure generation, and candidate evaluation are coordinated within one workflow [252]. MechAgents uses question–answer pairs extracted from source materials to construct a mechanical domain model for process simulation and behavior prediction, and has been applied to tasks such as composite-property prediction and stress–strain field analysis [253]. An integrated Mechanical Design Agent has also been developed for CAD design of joints and bolts, indicating the potential use of LLMs in mechanical design workflows [254].

More recently, LLM agents have also been coupled with computational mechanics modules. One example links a pre-trained LLM with a finite-element method (FEM) module to generate and optimize mechanical designs from natural-language specifications; in truss-structure design tasks, the reported success rate can reach 90% under specific constraints [255]. Related ideas have been explored in autonomous scientific experimentation. A data- and intelligence-driven machine chemist platform integrates a knowledge-processing module, robotic experimentation, and intelligent workstations, and was reported to synthesize a high-entropy catalyst within five weeks, compared with an estimated 1400 years by traditional trial-and-error methods [256]. This example shows how literature mining, knowledge-graph construction, experimental design, protocol optimization, and human–computer interaction can be combined within an automated research workflow [256]. High-throughput multi-agent systems have further been explored for materials research. MatPilot, for example, combines human expertise with AI-based knowledge abstraction, multimodal information processing, predictive modeling, optimization algorithms, and automated experimentation [257]. These systems can generate hypotheses, formulate experimental plans, invoke property-prediction models, and update subsequent experiments on the basis of feedback. For mechanical metastructures, such functions are useful because many systems still rely on manually designed algorithms or fixed models to realize specific unconventional responses. Changes in service environment, coupled physical fields, and task requirements often require repeated parameter adjustment and failure correction. Compared with conventional deep-learning models trained for narrow input–output mappings, LLM-based agents provide a practical interface for task decomposition, tool invocation, model coordination, and intermediate-result interpretation. Similar ideas have been applied to programmable metastructures, including an electromagnetic metaAgent that integrates natural-language processing with electromagnetic wave manipulation and human–robot interaction through a closed perception–decision–action loop [258].

As shown in Fig. 5, collaborative agents in a metastructure research framework can be organized through a central scheduling mechanism. The framework contains three main information flows. First, a multi-source and multimodal database provide literature data, high-throughput experimental results, process parameters, microstructural characterization, simulation outputs, defect and topology datasets, service-performance data, and application data for multi-scale components [227]. These data support structural feature engineering, multimodal data fusion, and intelligent analysis for candidate screening and knowledge extraction [229]. Second, domain agents and computational agents connect expert knowledge with simulation and design tasks. Manufacturing-process recommendation agents can be constructed for target–performance–guided multi-scale structural design, fault diagnosis on pilot platforms, online monitoring of manufacturing defects and residual stresses, process optimization, and lifecycle operation and maintenance [205,235]. High-throughput intelligent computing agents can further support forward performance prediction, property-driven inverse design, and optimization of metastructure primitives, unit cells, spatial topology, and process windows [80,241]. Their outputs should be constrained by mechanics-based models, manufacturing feasibility, uncertainty estimation, and validation data. Third, experimental agents provide physical feedback for model correction and design refinement. High-throughput experiments, performance verification tests, short-process manufacturing trials, and pilot-scale platforms can be used to evaluate candidate metastructures and update the computational workflow [230]. Multimodal robotic facility clusters may further support experimental monitoring, platform management, and automated decision-making in high-throughput metastructure experimentation [230]. Flexible intelligent manufacturing platforms can also integrate process-parameter monitoring, defect characterization, residual-stress evaluation, performance prediction, and spatial-topology analysis.

At the model layer, multimodal foundation-model platforms can be used for prediction, generation, control, and decision support. Possible components include expert-knowledge-based prediction models, integrated computational engineering models for composite materials and structures, damage-characterization models combined with in situ mechanical testing, multi-scale digital twin models, generative models for structural primitives and unit cells, geometric models for multi-scale components, constitutive models for metastructures, and multiphysics models for performance prediction and inverse design under extreme service conditions. With centralized data management and agent scheduling, the multimodal database can be updated continuously, and process–structure–performance relationships can be analyzed with improved consistency [253].

In summary, LLM-based multi-agent systems provide a possible workflow for linking literature agents, data agents, simulation agents, design agents, experimental agents, and pilot-platform agents in metastructure research. Their main role is to organize expert judgment, model execution, experimental feedback, and service information within a traceable research process. For LM³, this approach may help connect structure design, process optimization, performance verification, and application feedback, provided that the generated results remain constrained by mechanics and validated by simulation and experiment.

4.4 Mechanical metastructures and mechanoengineering

In 1959, Richard Feynman gave his well-known lecture “There’s Plenty of Room at the Bottom”, which is often regarded as an early conceptual origin of nanoscience and nanotechnology [259]. The central message of that lecture was not only that small scales were technically accessible, but also that new physical questions would emerge once matter could be manipulated with sufficient precision. Since then, nanoscale science and technology have reshaped information technology, instrumentation, manufacturing, and many aspects of daily life [260]. For mechanical metastructures, a related issue arises at larger engineering scales, where materials, structures, devices, and service systems must be designed in a coupled manner rather than treated as separate problems.

A useful analogy can be drawn from the development of biomechanics and mechanobiology. Classical biomechanics mainly studies motion, deformation, and force transmission in living systems [261,262]. It provides a mechanical description of biological structures and has long supported medical and health-related applications. Modern biomechanics further emphasizes quantitative analysis of mechanical factors in biological systems [263], where the multi-scale organization of cells, tissues, biological materials, and organs strongly affects both function and mechanical response [179,191,192]. However, this line of work is still often organized around forward analysis: given a biological structure or a loading condition, the task is to predict deformation, stress, or function. It is less naturally suited to design questions, such as how to create a material, structure, device, or treatment strategy to produce a prescribed biological or mechanical outcome [264,265].

Mechanobiology developed partly to address this limitation [261,264,265]. Rather than treating mechanics only as an external loading condition, it focuses on how mechanical cues influence growth, remodeling, adaptation, repair, disease progression, and cellular behavior [264-267]. The key point is feedback: stresses and strains affect tissue and cell states, while changes in these states reshape the local mechanical environment in return [261,266,267]. Mechanical factors, biological responses, and tissue or cellular remodeling are therefore coupled rather than arranged in a simple linear sequence. This perspective emphasizes the mutual dependence between structure and function. It has become useful not only for understanding normal growth, development, and aging, but also for diagnosis, intervention, device design, and therapeutic development [264-267].

A similar change of viewpoint is needed for mechanical metastructures. Conventional engineering design often separates material selection, structural analysis, manufacturing, and service assessment. Material properties are taken from handbooks, structural forms are designed within standard codes, and validation is carried out after the design has largely been fixed. This approach remains useful for many mature engineering systems, but it becomes restrictive when structures are expected to be lightweight, multi-functional, manufacturable, damage tolerant, and adaptive to complex service conditions [205-209,215,216,223]. Mechanical metastructures make this separation more difficult, because their performance is governed by topology, hierarchy, unit-cell geometry, defects, processing history, and boundary conditions at the same time [30-33,42-46,151,187]. Table 5 summarizes these mechanics-related concepts and clarifies how they contribute to LM³.

For LM³, the relevant design space is located not only at the material scale, but also at the levels of components, assemblies, equipment, and service systems. In this design space, architecture, mechanics, manufacturing, sensing, and control are coupled [208-210,216,217,227,238]. Recent developments in high-throughput experiments [223,225,230-233], high-throughput simulations [221,222,230], digital twins, artificial intelligence [241-258], and the Structural Genome Initiative [216,217,224-226] provide methods for exploring this space. However, these methods do not by themselves constitute a theory. They need to be connected through mechanical principles that explain how structural genomes, spatial topology, fabrication processes, and service environments jointly determine performance.

Mechanoengineering is used here to describe this broader design perspective. It extends beyond the analysis of isolated materials or conventional structural elements, and focuses on the co-design of materials, architectures, manufacturing routes, and service performance at component and system levels [272,273]. For LM³, mechanoengineering provides a way to organize several questions that otherwise remain scattered: how to define structural genomes, how to transfer properties across scales, how to incorporate manufacturing defects, how to combine load-bearing and functional requirements, and how to close the loop among simulation, fabrication, testing, and service feedback. As depicted in Fig. 6, mechanical metastructures can therefore be positioned within a mechanoengineering paradigm that connects fundamental mechanics, structural-genome design, high-throughput experimentation, intelligent computation, and industrial application [208,216,217,223,225,241-258,271-273].

4.5 Four-beam and eight-column framework for AI4MechanoStructure

As illustrated in Fig. 7, the AI4MechanoStructure paradigm can be organized through a four-beam and eight-column framework. The four beams represent the core capability platforms required for the research, development, and deployment of mechanical metastructures, whereas the eight columns correspond to representative industrial application domains. The four capability platforms include the mechanical metastructure database platform [226,240], the intelligent computing platform [228,237], the intelligent experiment platform [230,235], and the mechanical metastructure large-model platform [252,253]. Rather than functioning as isolated modules, these platforms support an integrated research workflow in which data, models, experiments, and engineering decisions are exchanged during design, fabrication, testing, qualification, and service assessment.

The database platform provides the data infrastructure for this framework. It collects, organizes, and manages multi-scale and multimodal information associated with structural genomes, unit-cell topologies, manufacturing parameters, defect characteristics, mechanical responses, and service performance [226]. Such a platform is necessary because metastructure-related information is often distributed across simulations, experiments, imaging data, process records, and application-specific evaluations. A unified database improves data traceability and reproducibility, facilitates comparisons among different studies, and supplies training datasets for property prediction, inverse design, uncertainty quantification, and process optimization.

The intelligent computing platform supports modeling, simulation, and design activities within the framework [237]. It integrates multi-scale simulation, topology optimization, surrogate modeling, uncertainty analysis, and multi-objective optimization to screen metastructures with prescribed stiffness, strength, energy absorption, thermal regulation, acoustic performance, or other functional characteristics. Compared with conventional trial-and-error approaches, intelligent computing allows candidate architectures to be evaluated before fabrication, while retaining compatibility with mechanics-based constraints and manufacturing requirements [228,237].

The intelligent experiment platform provides the physical validation and feedback mechanisms required by the framework [230,235]. Through robotic automation, in situ characterization, high-throughput fabrication, and high-throughput testing, it can shorten empirical development cycles while generating experimental data for model calibration and verification. This capability is particularly important for mechanical metastructures because manufacturing deviations, surface conditions, internal defects, and boundary effects can substantially influence measured performance. By feeding experimental observations back into the database and computing platforms, the intelligent experiment platform helps establish a closed-loop workflow linking virtual design and physical realization [230,235].

The large-model platform serves as an interface for knowledge extraction, design assistance, reasoning, and human–AI collaboration [252,253]. Domain-adapted foundation models can assist researchers in literature retrieval, knowledge synthesis, design-rule extraction, candidate concept generation, code development, simulation workflow management, and natural-language interaction. Large models should not be regarded as autonomous design engines. Their outputs must remain constrained by mechanical principles, validated through simulation, and verified experimentally. When integrated with the database, computing, and experiment platforms, the large-model platform can support data-driven prediction, inverse design, process planning, and service-oriented optimization.

The eight columns represent the principal application domains supported by these four capability platforms. Representative sectors include aerospace engineering, civil engineering, transportation engineering, smart electronics, ocean engineering, biomedical engineering, defense engineering, and textile engineering. Although these domains differ substantially in operating environments and performance requirements, they share several common objectives, including lightweight design, structural reliability, functional integration, manufacturability, and adaptability to complex service conditions. The four-beam framework therefore provides the enabling research infrastructure, whereas the eight columns provide application-driven requirements and feedback. Engineering demands guide the selection of structural genomes, materials, architectures, and manufacturing routes, while advances in databases, intelligent computation, automated experimentation, and large models increase the number of metastructure solutions that can be evaluated for practical deployment.

Overall, the four-beam and eight-column framework should be viewed as an organizational architecture for AI-enabled mechanical metastructure research and development. It connects data infrastructure, intelligent computation, automated experimentation, and foundation-model assistance with the requirements of major engineering sectors. For LM³, such a framework provides a way to move from isolated structural concepts toward validated, manufacturable, and service-ready mechanical metastructure systems.

5 Conclusions and Outlooks

Lightweight multi-functional mechanical metastructures should be understood as architecture-governed structural systems rather than as simple extensions of lightweight materials or periodic lattices. Their performance is determined by the combined effects of constituent materials, unit-cell topology, hierarchy, connectivity, manufacturing quality, defects, boundary conditions, and service environments. This review has shown that structural genomes, multi-scale modeling, topology optimization, high-throughput experimentation, and digital twin methods provide a basis for describing and designing such systems. These approaches make it possible to pursue property combinations that are difficult to achieve in conventional materials or monolithic structures, including lightweight load bearing, energy absorption, vibration regulation, thermal management, sensing, and adaptive response.

Artificial intelligence can further support LM³ design when it is coupled with mechanics-based models, fabrication constraints, and experimental validation. Data-driven prediction, physics-informed learning, generative and inverse design, bioinspired strategies, SGI frameworks, and LLM-based multi-agent systems can improve the efficiency of structure screening, process optimization, and closed-loop experimentation. At the same time, their reliability depends on standardized structural-genome representations, traceable multimodal datasets, interpretable models, robust manufacturing routes, defect-aware characterization, and service-oriented testing. Future work should therefore move beyond isolated unit-cell demonstrations and focus on validated, manufacturable, and application-specific LM³ systems within the broader mechanoengineering and AI4MechanoStructure frameworks.

Future progress in lightweight multifunctional mechanical metastructures will depend on a transition from topology-centered proof-of-concept studies toward lifecycle-oriented and quantitatively validated engineering systems. A primary scientific challenge is to establish a standardized structural-genome representation that consistently describes topology, connectivity, hierarchy, constituent materials, manufacturing parameters, geometric deviations, defects, loading histories, and service environments. On this basis, FAIR and multimodal benchmark datasets integrating simulations, in situ characterization, manufacturing records, and component-level tests will be essential for reliable model development and cross-study comparison. Future AI models should therefore move beyond purely data-driven correlations toward physics-informed, uncertainty-aware, and defect-sensitive frameworks. In addition to prediction accuracy, greater attention should be paid to extrapolation capability, robustness under distribution shifts, interpretability, and the propagation of uncertainty across multiple structural scales.

The convergence of digital twins, the Structural Genome Initiative, high-throughput simulation and experimentation, intelligent manufacturing, and LLM-enabled multi-agent systems may ultimately create a closed-loop workflow linking design, fabrication, characterization, validation, and service feedback. In such a framework, AI agents can assist with knowledge retrieval, candidate generation, simulation execution, process optimization, and experimental planning, while mechanics-based constraints and human expert supervision ensure physical validity and traceability. Importantly, future metastructures should be evaluated not only by their idealized stiffness, strength, or multifunctionality, but also by manufacturability, defect tolerance, fatigue durability, scalability, sustainability, cost, and reliability under realistic service conditions. Addressing these issues will enable mechanical metastructures to evolve from isolated architected concepts into integrated mechanoengineering platforms for aerospace, transportation, robotics, biomedical devices, and other advanced engineering systems. In this sense, there remains considerable room at the top—not merely for discovering new architectures, but for transforming architectural mechanics into dependable and adaptive engineering technologies.

References

[1] Jiao P C, Alavi A H. Artificial intelligence-enabled smart mechanical metamaterials: advent and future trends. International Materials Reviews, 2021, 66(6): 365–393

[2] Challapalli A, Patel D, Li G Q. Inverse machine learning framework for optimizing lightweight metamaterials. Materials & Design, 2021, 208: 109937

[3] Mao H N, Rumpler R, Gaborit M. et al. Twist, tilt and stretch: from isometric Kelvin cells to anisotropic cellular materials. Materials & Design, 2020, 193: 108855

[4] Cui Z M, Ju J. Mechanical coupling effects of 2D lattices uncovered by decoupled micropolar elasticity tensor and symmetry operation. Journal of the Mechanics and Physics of Solids, 2022, 167: 105012

[5] Rodríguez S E, Calius E P, Khatibi A. et al. Mechanical metamaterial systems as transformation mechanisms. Extreme Mechanics Letters, 2023, 61: 101985

[6] Chen Y Y, Nassar H, Huang G L. Discrete transformation elasticity: an approach to design lattice-based polar metamaterials. International Journal of Engineering Science, 2021, 168: 103562

[7] Sutton G P, Biblarz O. Rocket Propulsion Elements. 9th ed. New York: John Wiley & Sons, 2016

[8] Jozič P, Zidanšek A, Repnik R. Fuel conservation for launch vehicles: falcon heavy case study. Energies, 2020, 13(3): 660

[9] Fahy W P, Langston J, Wu H. et al. Silica-phenolic nanocomposite ablatives for thermal protection application. Journal of Spacecraft and Rockets, 2020, 57(3): 596–602

[10] Li D, Li P Q. Technological breakthroughs of LM-5 and future developments of China’s launch vehicle. Acta Aeronautica et Astronautica Sinica, 2022, 43(10): 527269

[11] Smeets B, Pavlov L, Kassapoglou C. Development and testing of equipment attachment zones for lattice and grid-stiffened composite structures. In: Proceedings of the 14th European Conference on Spacecraft Structures, Materials and Environmental Testing. Toulouse, France, 2016 –57490

[12] Singh A, Al-Ketan O, Karathanasopoulos N. Hybrid manufacturing of AlSi10Mg metamaterials: process, static and impact response attributes. Journal of Materials Research and Technology, 2023, 27: 7457–7469

[13] Li J L, Zhao J, Sun Z. et al. Lightweight design of transmission frame structures for launch vehicles based on moving morphable components (MMC) approach. Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(1): 244–251

[14] Totaro G, Spena P, Giusto G. et al. Highly efficient CFRP anisogrid lattice structures for central tubes of medium-class satellites: design, manufacturing, and performance. Composite Structures, 2021, 258: 113368

[15] Xue T J, Adriaenssens S, Mao S. Learning the nonlinear dynamics of mechanical metamaterials with graph networks. International Journal of Mechanical Sciences, 2023, 238: 107835

[16] Hernández Q, Badías A, González D. et al. Structure-preserving neural networks. Journal of Computational Physics, 2021, 426: 109950

[17] Liao W H, Dai N. Development and challenge of lightweight design and manufacturing technology for aerospace structures. Journal of Nanjing University of Aeronautics & Astronautics, 2023, 55(3): 347–360

[18] Bravo-Mosquera P D, Catalano F M, Zingg D W. Unconventional aircraft for civil aviation: a review of concepts and design methodologies. Progress in Aerospace Sciences, 2022, 131: 100813

[19] Lee D, Chen W, Wang L W. et al. Data-driven design for metamaterials and multi-scale systems: a review. Advanced Materials, 2024, 36(8): 2305254

[20] Zhao Z Y, Li L, Wang X. et al. Strength optimization of ultralight corrugated-channel-core sandwich panels. Science China Technological Sciences, 2019, 62(8): 1467–1477

[21] Chu Z Q, Wang R C, Tian S B. et al. Fabrication and failure mechanisms of ultralight all-CFRP sandwich cylinders under axial compression. Composite Structures, 2024, 345: 118386

[22] Cao M H, Yan H B, Xiao P F. et al. Multifunctional design of an X-lattice interlocked sandwich structure with integrated electromagnetic wave regulation, convective heat transfer and load bearing performances. Composite Structures, 2024, 345: 118401

[23] Qu Z G, Wang T S, Tao W Q. et al. Experimental study of air natural convection on metallic foam-sintered plate. International Journal of Heat and Fluid Flow, 2012, 38: 126–132

[24] Woods W D. Earth orbit and TLI. In: Woods W D, ed. How Apollo Flew to the Moon. New York: Springer, 2011 107–130

[25] Li L B, Guo Z M, Yang F. et al. Additively manufactured acoustic-mechanical multifunctional hybrid lattice structures. International Journal of Mechanical Sciences, 2024, 269: 109071

[26] Li X W, Yu X, Zhai W. Additively manufactured deformation-recoverable and broadband sound-absorbing microlattice inspired by the concept of traditional perforated panels. Advanced Materials, 2021, 33(44): 2104552

[27] Li X W, Yu X, Chua J W. et al. Microlattice metamaterials with simultaneous superior acoustic and mechanical energy absorption. Small, 2021, 17(24): 2100336

[28] Li Z D, Li X W, Wang Z G. et al. Multifunctional sound-absorbing and mechanical metamaterials via a decoupled mechanism design approach. Materials Horizons, 2023, 10(1): 75–87

[29] Zhang X, Yao J H, Liu B. et al. Three-dimensional high-entropy alloy-polymer composite nanolattices that overcome the strength-recoverability trade-off. Nano Letters, 2018, 18(7): 4247–4256

[30] Fleck N A, Deshpande V S, Ashby M F. Micro-architectured materials: past, present and future. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2010, 466(2121): 2495–2516

[31] Zhong H Z, Song T T, Li C W. et al. The Gibson-Ashby model for additively manufactured metal lattice materials: its theoretical basis, limitations and new insights from remedies. Current Opinion in Solid State and Materials Science, 2023, 27(3): 101081

[32] Alomar Z, Concli F. A review of the selective laser melting lattice structures and their numerical models. Advanced Engineering Materials, 2020, 22(12): 2000611

[33] Wu X X, Yan H, Zhou Y Q. et al. Review of additive manufactured metallic metamaterials: design, fabrication, property and application. Optics & Laser Technology, 2025, 182: 112066

[34] Wang B, Tian K, Zhou C H. et al. Grid-pattern optimization framework of novel hierarchical stiffened shells allowing for imperfection sensitivity. Aerospace Science and Technology, 2017, 62: 114–121

[35] Zhang S P, Guo H M, Gao T. et al. Design and manufacturing method of multi-scale integrated load bearing thin-walled structure for application in next-generation aeroengine based on advanced laser processing technology. Acta Aeronautica et Astronautica Sinica, 2024, 45(13): 630037

[36] Zhang X Y, Liu C, Shi L M. et al. Optimal design of shell-lattice infill integrated supporting structure based on the method of moving morphable components and its application in China Space Station. Chinese Journal of Solid Mechanics, 2022, 43(5): 551–563

[37] Yue Z S, Wang X, He C. et al. Elevated shock resistance of all-metallic sandwich beams with honeycomb-supported corrugated cores. Composites Part B: Engineering, 2022, 242: 110102

[38] Wang C, Zhu J H, Wu M Q. et al. Multi-scale design and optimization for solid-lattice hybrid structures and their application to aerospace vehicle components. Chinese Journal of Aeronautics, 2021, 34(5): 386–398

[39] Jeong H S, Lyu S K, Park S H. Effective strut-based design approach of multi-shaped lattices using equivalent material properties. Journal of Mechanical Science and Technology, 2021, 35(4): 1609–1622

[40] Wang X, Qin R X, Chen B Z. et al. Multi-scale collaborative optimization of lattice structures using laser additive manufacturing. International Journal of Mechanical Sciences, 2022, 222: 107257

[41] nTop. Aerojet Rocketdyne’s 3D-printed quad thruster enables low-cost space exploration. Available online (accessed 09 Apr 2026)

[42] Xiao M, Liu X L, Zhang Y. et al. Design of graded lattice sandwich structures by multi-scale topology optimization. Computer Methods in Applied Mechanics and Engineering, 2021, 384: 113949

[43] Zhang Y, Xiao M, Ding Z. et al. Dynamic response-oriented multi-scale topology optimization for geometrically asymmetric sandwich structures with graded cellular cores. Computer Methods in Applied Mechanics and Engineering, 2023, 416: 116367

[44] Zhang X Y, Zhou H, Shi W H. et al. Vibration tests of 3D printed satellite structure made of lattice sandwich panels. AIAA Journal, 2018, 56(10): 4213–4217

[45] Liu Q Y, Xu R W, Zhou Y. et al. Metamaterials mapped lightweight structures by principal stress lines and topology optimization: methodology, additive manufacturing, ductile failure and tests. Materials & Design, 2021, 212: 110192

[46] Jia Y Q, Liu K, Zhang X S. Topology optimization of irregular multi-scale structures with tunable responses using a virtual growth rule. Computer Methods in Applied Mechanics and Engineering, 2024, 425: 116864

[47] Jiang Y F, Shen C, Meng H. et al. Design and optimization of micro-perforated ultralight sandwich structure with N-type hybrid core for broadband sound absorption. Applied Acoustics, 2023, 202: 109184

[48] Jiang Y F, Feng X C, Gao J L. et al. Design of ultralight multifunctional sandwich structure with n-h hybrid core for integrated sound absorption and load-bearing capacity. Materials Today Communications, 2024, 41: 110663

[49] Rahman M, Fricks C, Ahmed H. et al. Topology optimization and experimental validation of mass-reduced aircraft wing designs. Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 2024, 239(4): 326–337

[50] Jia H R, Duan S Y, Zhang Z. et al. Homogenization-based topology optimization for self-supporting additive-manufactured lattice-infilled structure. Materials & Design, 2024, 245: 113264

[51] Liu Q Y, Zhou Y, Zhang Z J. et al. Crystal sheet lattices: novel mechanical metamaterials with smooth profiles, reduced anisotropy, and enhanced mechanical performances. Materials & Design, 2022, 223: 111123

[52] Milton G W. Composite materials with Poisson's ratios close to −1. Journal of the Mechanics and Physics of Solids, 1992, 40(5): 1105–1137

[53] Christensen J, Kadic M, Wegener M. et al. Vibrant times for mechanical metamaterials. MRS Communications, 2015, 5(3): 453–462

[54] Crook C, Bauer J, Izard A G. et al. Plate-nanolattices at the theoretical limit of stiffness and strength. Nature Communications, 2020, 11(1): 1579

[55] Zhang L, Liu H W, Song B. et al. Wood-inspired metamaterial catalyst for robust and high-throughput water purification. Nature Communications, 2024, 15(1): 2046

[56] Huang W, Shishehbor M, Guarín-Zapata N. et al. A natural impact-resistant bicontinuous composite nanoparticle coating. Nature Materials, 2020, 19(11): 1236–1243

[57] Dong L, Lakes R. Advanced damper with high stiffness and high hysteresis damping based on negative structural stiffness. International Journal of Solids and Structures, 2013, 50(14-15): 2416–2423

[58] Hou Y Z, Guan Q F, Xia J. et al. Strengthening and toughening hierarchical nanocellulose via humidity-mediated interface. ACS Nano, 2021, 15(1): 1310–1320

[59] Sajadi S M, Vásárhelyi L, Mousavi R. et al. Damage-tolerant 3D-printed ceramics via conformal coating. Science Advances, 2021, 7(28): eabc5028

[60] Chen Z Y, Xie Y M, Wu X. et al. On hybrid cellular materials based on triply periodic minimal surfaces with extreme mechanical properties. Materials & Design, 2019, 183: 108109

[61] Noronha J, Dash J, Rogers J. et al. Titanium multi-topology metamaterials with exceptional strength. Advanced Materials, 2024, 36(34): 2308715

[62] Guo X Y, Li W B, Zhang W M. Adjustable stiffness elastic composite soft actuator for fast-moving robots. Science China Technological Sciences, 2021, 64(8): 1663–1675

[63] Ji X B, Liu X C, Cacucciolo V. et al. An autonomous untethered fast soft robotic insect driven by low-voltage dielectric elastomer actuators. Science Robotics, 2019, 4(37): eaaz6451

[64] Gao Y, Wu Q Q, Wei X Y. et al. Composite tree-like re-entrant structure with high stiffness and controllable elastic anisotropy. International Journal of Solids and Structures, 2020, 206: 170–182

[65] Wang D H, Sun Q Q, Hokkanen M J. et al. Design of robust superhydrophobic surfaces. Nature, 2020, 582(7810): 55–59

[66] Li C X, Yu C L, Zhou S. et al. Liquid harvesting and transport on multi-scaled curvatures. Proceedings of the National Academy of Sciences of the United States of America, 2020, 117(38): 23436–23442

[67] Berger J, Mercer C, McMeeking R M. et al. The design of bonded bimaterial lattices that combine low thermal expansion with high stiffness. Journal of the American Ceramic Society, 2011, 94(s1): s42–s54

[68] Kalnaus S, Asp L E, Li J L. et al. Multifunctional approaches for safe structural batteries. Journal of Energy Storage, 2021, 40: 102747

[69] Liu Y D, Yin Y H, Guo Z Z. Static and dynamic design based on hierarchical optimization for materials and structure of porous metals. Science China Technological Sciences, 2012, 55(10): 2808–2814

[70] Yang W. Digintel mechanics—Governing the digintel era. Advances in Mechanics, 2024, 54(4): 629–638

[71] Huang E W, Lee W J, Singh S S. et al. Machine-learning and high-throughput studies for high-entropy materials. Materials Science and Engineering: R: Reports, 2022, 147: 100645

[72] Ulloa J, Ariza M P, Andrade J E. et al. Homogenized models of mechanical metamaterials. Computer Methods in Applied Mechanics and Engineering, 2025, 433: 117454

[73] Bordiga G, Medina E, Jafarzadeh S. et al. Automated discovery of reprogrammable nonlinear dynamic metamaterials. Nature Materials, 2024, 23(11): 1486–1494

[74] Liu Z L, Wu C T. Exploring the 3D architectures of deep material network in data-driven multi-scale mechanics. Journal of the Mechanics and Physics of Solids, 2019, 127: 20–46

[75] Bessa M A, Glowacki P, Houlder M. Bayesian machine learning in metamaterial design: fragile becomes supercompressible. Advanced Materials, 2019, 31(48): 1904845

[76] Yang C, Kim Y, Ryu S. et al. Prediction of composite microstructure stress-strain curves using convolutional neural networks. Materials & Design, 2020, 189: 108509

[77] Yang Z Z, Yu C H, Buehler M J. Deep learning model to predict complex stress and strain fields in hierarchical composites. Science Advances, 2021, 7(15): eabd7416

[78] Chai Z P, Zong Z S, Yong H C. et al. Tailoring stress-strain curves of flexible snapping mechanical metamaterial for on-demand mechanical responses via data-driven inverse design. Advanced Materials, 2024, 36(33): 2404369

[79] Bastek J H, Kumar S, Telgen B. et al. Inverting the structure-property map of truss metamaterials by deep learning. Proceedings of the National Academy of Sciences of the United States of America, 2022, 119(1): e2111505119

[80] Ha C S, Yao D S, Xu Z P. et al. Rapid inverse design of metamaterials based on prescribed mechanical behavior through machine learning. Nature Communications, 2023, 14(1): 5765

[81] Gu G X, Wettermark S, Buehler M J. Algorithm-driven design of fracture resistant composite materials realized through additive manufacturing. Additive Manufacturing, 2017, 17: 47–54

[82] Lee S, Zhang Z Z, Gu G X. Generative machine learning algorithm for lattice structures with superior mechanical properties. Materials Horizons, 2022, 9(3): 952–960

[83] Mao Y W, He Q, Zhao X H. Designing complex architectured materials with generative adversarial networks. Science Advances, 2020, 6(17): eaaz4169

[84] Oh S, Jung Y, Kim S. et al. Deep generative design: Integration of topology optimization and generative models. Journal of Mechanical Design, 2019, 141(11): 111405

[85] Javadi A A, Faramarzi A, Farmani R. Design and optimization of microstructure of auxetic materials. Engineering Computations, 2012, 29(3): 260–276

[86] Abdeljaber O, Avci O, Inman D J. Optimization of chiral lattice based metastructures for broadband vibration suppression using genetic algorithms. Journal of Sound and Vibration, 2016, 369: 50–62

[87] Yeh S L, Harne R L. Origins of broadband vibration attenuation empowered by optimized viscoelastic metamaterial inclusions. Journal of Sound and Vibration, 2019, 458: 218–237

[88] Pokkalla D K, Poh L H, Quek S T. Isogeometric shape optimization of missing rib auxetics with prescribed negative Poisson's ratio over large strains using genetic algorithm. International Journal of Mechanical Sciences, 2021, 193: 106169

[89] Wang L, Liu H T. Parameter optimization of bidirectional re-entrant auxetic honeycomb metamaterial based on genetic algorithm. Composite Structures, 2021, 267: 113915

[90] Vangelatos Z, Sheikh H M, Marcus P S. et al. Strength through defects: a novel Bayesian approach for the optimization of architected materials. Science Advances, 2021, 7(41): eabk2218

[91] Vaissier B, Pernot J P, Chougrani L. et al. Genetic-algorithm based framework for lattice support structure optimization in additive manufacturing. Computer-Aided Design, 2019, 110: 11–23

[92] Gao Z Y, Zhang X L, Wu Y. et al. Damage-programmable design of metamaterials achieving crack-resisting mechanisms seen in nature. Nature Communications, 2024, 15(1): 7373

[93] Tran A, Tran M, Wang Y. Constrained mixed-integer Gaussian mixture Bayesian optimization and its applications in designing fractal and auxetic metamaterials. Structural and Multidisciplinary Optimization, 2019, 59(6): 2131–2154

[94] Bessa M A, Pellegrino S. Design of ultra-thin shell structures in the stochastic post-buckling range using Bayesian machine learning and optimization. International Journal of Solids and Structures, 2018, 139–140: 174–188

[95] Isanaka B R, Mukhopadhyay T, Varma R K. et al. On exploiting machine learning for failure pattern driven strength enhancement of honeycomb lattices. Acta Materialia, 2022, 239: 118226

[96] Shi K Y, Gu D D, Liu H. et al. Process-structure multi-objective inverse optimisation for additive manufacturing of lattice structures using a physics-enhanced data-driven method. Virtual and Physical Prototyping, 2023, 18(1): e2266641

[97] Qi G, Ma L, Wang S Y. Modeling and reliability of insert in composite pyramidal lattice truss core sandwich panels. Composite Structures, 2019, 221: 110888

[98] Konstantopoulos G, Koumoulos E P, Charitidis C A. Classification of mechanism of reinforcement in the fiber-matrix interface: application of machine learning on nanoindentation data. Materials & Design, 2020, 192: 108705

[99] Morsali S, Qian D, Minary-Jolandan M. Designing bioinspired brick-and-mortar composites using machine learning and statistical learning. Communications Materials, 2020, 1(1): 12

[100] Zhong Z Y, An J, Wu D. et al. A machine learning strategy for enhancing the strength and toughness in metal matrix composites. International Journal of Mechanical Sciences, 2024, 281: 109550

[101] Lan X K, Huang Q, Zhou T. et al. Optimal design of a novel cylindrical sandwich panel with double arrow auxetic core under air blast loading. Defence Technology, 2020, 16(3): 617–626

[102] Alwattar T A, Mian A. Development of an elastic material model for BCC lattice cell structures using finite element analysis and neural networks approaches. Journal of Composites Science, 2019, 3(2): 33

[103] Wu L L, Liu L, Wang Y. et al. A machine learning-based method to design modular metamaterials. Extreme Mechanics Letters, 2020, 36: 100657

[104] Kumar S, Tan S, Zheng L. et al. Inverse-designed spinodoid metamaterials. npj Computational Materials, 2020, 6(1): 73

[105] Kulagin R, Beygelzimer Y, Estrin Y. et al. Architectured lattice materials with tunable anisotropy: design and analysis of the material property space with the aid of machine learning. Advanced Engineering Materials, 2020, 22(12): 2001069

[106] Pahlavani H, Amani M, Cruz Saldívar M. et al. Deep learning for the rare-event rational design of 3D printed multi-material mechanical metamaterials. Communications Materials, 2022, 3(1): 46

[107] Wang Y Z, Zeng Q L, Wang J Z. et al. Inverse design of shell-based mechanical metamaterial with customized loading curves based on machine learning and genetic algorithm. Computer Methods in Applied Mechanics and Engineering, 2022, 401: 115571

[108] Lee S, Zhang Z Z, Gu G X. Deep learning accelerated design of mechanically efficient architected materials. ACS Applied Materials & Interfaces, 2023, 15(18): 22543–22552

[109] Maurizi M, Gao C, Berto F. Inverse design of truss lattice materials with superior buckling resistance. npj Computational Materials, 2022, 8(1): 247

[110] Muhammad W, Brahme A P, Ibragimova O. et al. A machine learning framework to predict local strain distribution and the evolution of plastic anisotropy & fracture in additively manufactured alloys. International Journal of Plasticity, 2021, 136: 102867

[111] Peng X L, Xu B X. Data-driven inverse design of composite triangular lattice structures. International Journal of Mechanical Sciences, 2024, 265: 108900

[112] Meyer P P, Tancogne-Dejean T, Mohr D. Non-symmetric plate-lattices: Recurrent neural network-based design of optimal metamaterials. Acta Materialia, 2024, 278: 120246

[113] Shendy M, Alkhader M, Abu-Nabah B A. et al. Machine learning assisted approach to design lattice materials with prescribed band gap characteristics. European Journal of Mechanics - A/Solids, 2023, 102: 105125

[114] Ma C P, Zhang Z W, Luce B. et al. Accelerated design and characterization of non-uniform cellular materials via a machine-learning based framework. npj Computational Materials, 2020, 6(1): 40

[115] Hassanin H, Alkendi Y, Elsayed M. et al. Controlling the properties of additively manufactured cellular structures using machine learning approaches. Advanced Engineering Materials, 2020, 22(3): 1901338

[116] Teimouri A, Challapalli A, Konlan J. et al. Machine learning assisted design and optimization of plate-lattice structures with superior specific recovery force. Giant, 2024, 18: 100282

[117] Yue Z S, Han B, Wang Z Y. et al. Data-driven multi-objective optimization of ultralight hierarchical origami-corrugation meta-sandwich structures. Composite Structures, 2023, 303: 116334

[118] Jia Y Q, Liu K, Zhang X S. Modulate stress distribution with bio-inspired irregular architected materials towards optimal tissue support. Nature Communications, 2024, 15(1): 4072

[119] Gao Z Y, Wang H Z, Letov N. et al. Data-driven design of biometric composite metamaterials with extremely recoverable and ultrahigh specific energy absorption. Composites Part B: Engineering, 2023, 251: 110468

[120] Chen R G, Zhang W J, Jia Y F. et al. Ultra-stiff and quasi-elastic-isotropic triply periodic minimal surface structures designed by deep learning. Materials & Design, 2024, 244: 113107

[121] Zhang K, Guo Y Y, Liu X B. et al. Deep learning-based inverse design of lattice metamaterials for tuning bandgap. Extreme Mechanics Letters, 2024, 69: 102165

[122] Jia Z B, Gong H, Liu S Y. et al. Designing three-dimensional lattice structures with anticipated properties through a deep learning method. Materials & Design, 2024, 244: 113139

[123] Wang Y J, Liao Z Y, Shi S Y. et al. Data-driven structural design optimization for petal-shaped auxetics using isogeometric analysis. Computer Modeling in Engineering & Sciences, 2020, 122(2): 433–458

[124] Ji Z K, Li D W, Zhang C D. et al. AI-aided design and multi-scale optimization of mechanical metastructures with controllable anisotropy. Engineering Structures, 2024, 310: 118134

[125] Wu Y R, Mao Z F, Feng Y Q. Energy absorption prediction for lattice structure based on D2 shape distribution and machine learning. Composite Structures, 2023, 319: 117136

[126] Peng C X, Tran P, Rutz E. Accelerating hybrid lattice structures design with machine learning. Materials Science in Additive Manufacturing, 2024, 3(2): 3430

[127] Matthews J, Klatt T, Morris C. et al. Hierarchical design of negative stiffness metamaterials using a Bayesian network classifier. Journal of Mechanical Design, 2016, 138(4): 041404

[128] Wilt J K, Yang C, Gu G X. Accelerating auxetic metamaterial design with deep learning. Advanced Engineering Materials, 2020, 22(5): 2070018

[129] Zhang Z Z, Gu G X. Finite-element-based deep-learning model for deformation behavior of digital materials. Advanced Theory and Simulations, 2020, 3(7): 2000031

[130] Liu T W, Sun S W, Liu H. et al. A predictive deep-learning approach for homogenization of auxetic kirigami metamaterials with randomly oriented cuts. Modern Physics Letters B, 2020, 35(1): 2150033

[131] Hanakata P Z, Cubuk E D, Campbell D K. et al. Accelerated search and design of stretchable graphene kirigami using machine learning. Physical Review Letters, 2018, 121(25): 255304

[132] Gu G X, Chen C T, Buehler M J. De novo composite design based on machine learning algorithm. Extreme Mechanics Letters, 2018, 18: 19–28

[133] Gu G X, Chen C T, Richmond D J. et al. Bioinspired hierarchical composite design using machine learning: Simulation, additive manufacturing, and experiment. Materials Horizons, 2018, 5(5): 939–945

[134] Yang Z J, Yabansu Y C, Jha D. et al. Establishing structure-property localization linkages for elastic deformation of three-dimensional high contrast composites using deep learning approaches. Acta Materialia, 2019, 166: 335–345

[135] Wei A R, Xiong J, Yang W D. et al. Deep learning-assisted elastic isotropy identification for architected materials. Extreme Mechanics Letters, 2021, 43: 101173

[136] Zhao T Y, Li Y W, Zuo L. et al. Machine-learning optimized method for regional control of sound fields. Extreme Mechanics Letters, 2021, 45: 101297

[137] Zhang J Y, Li Y W, Zhao T Y. et al. Machine-learning based design of digital materials for elastic wave control. Extreme Mechanics Letters, 2021, 48: 101372

[138] Peng B, Wei Y, Qin Y. et al. Machine learning-enabled constrained multi-objective design of architected materials. Nature Communications, 2023, 14(1): 6630

[139] Zhao M, Li X W, Yan X. et al. Machine learning accelerated design of lattice metamaterials for customizable energy absorption. Thin-Walled Structures, 2025, 208: 112845

[140] Garland A P, White B C, Jensen S C. et al. Pragmatic generative optimization of novel structural lattice metamaterials with machine learning. Materials & Design, 2021, 203: 109632

[141] Yang Z J, Yabansu Y C, Al-Bahrani R. et al. Deep learning approaches for mining structure-property linkages in high contrast composites from simulation datasets. Computational Materials Science, 2018, 151: 278–287

[142] Kollmann H T, Abueidda D W, Koric S. et al. Deep learning for topology optimization of 2D metamaterials. Materials & Design, 2020, 196: 109098

[143] Bonfanti S, Guerra R, Font-Clos F. et al. Automatic design of mechanical metamaterial actuators. Nature Communications, 2020, 11(1): 4162

[144] Garland A P, White B C, Jared B H. et al. Deep convolutional neural networks as a rapid screening tool for complex additively manufactured structures. Additive Manufacturing, 2020, 35: 101217

[145] Ma C P, Chang Y L, Wu S. et al. Deep learning-accelerated designs of tunable magneto-mechanical metamaterials. ACS Applied Materials & Interfaces, 2022, 14(29): 33892–33902

[146] Wang Z F, Wang S, Ma C W. et al. The prediction of homogenized effective properties of continuous fiber composites based on a deep transfer learning approach. Composites Science and Technology, 2025, 262: 111050

[147] Li Y S, Qin H S, Jia L Y. et al. Microstructure dependent transverse strength criterion for UD-CFRP composites via computational micromechanics and machine learning. Composites Science and Technology, 2024, 251: 110551

[148] Li S Y, Tian X X, Li Q B. et al. Advancing structural health monitoring: deep learning-enhanced quantitative analysis of damage in composite laminates using surface strain field. Composites Science and Technology, 2024, 258: 110880

[149] Yang H, Wang W F, Li C L. et al. Deep learning-based X-ray computed tomography image reconstruction and prediction of compression behavior of 3D printed lattice structures. Additive Manufacturing, 2022, 54: 102774

[150] Zhang P, Tang K K, Chen G X. et al. Multimodal data fusion enhanced deep learning prediction of crack path segmentation in CFRP composites. Composites Science and Technology, 2024, 257: 110812

[151] Meyer P P, Bonatti C, Tancogne-Dejean T. et al. Graph-based metamaterials: deep learning of structure-property relations. Materials & Design, 2022, 223: 111175

[152] Xiao L J, Shi G Q, Song W D. Machine learning predictions on the compressive stress-strain response of lattice-based metamaterials. International Journal of Solids and Structures, 2024, 300: 112893

[153] Zheng L, Karapiperis K, Kumar S. et al. Unifying the design space and optimizing linear and nonlinear truss metamaterials by generative modeling. Nature Communications, 2023, 14(1): 7563

[154] Qiu C, Han Y Z, Shanmugam L, et al. A deep learning-based composite design strategy for efficient selection of material and layup sequences from a given database. Composites Science and Technology, 2022, 230(Pt 2): 109154

[155] Yüksel N, Eren O, Börklü H R. et al. Mechanical properties of additively manufactured lattice structures designed by deep learning. Thin-Walled Structures, 2024, 196: 111475

[156] Yang Z Z, Yu C H, Guo K. et al. End-to-end deep learning method to predict complete strain and stress tensors for complex hierarchical composite microstructures. Journal of the Mechanics and Physics of Solids, 2021, 154: 104506

[157] Zhang H J, Wang Y, Zhao H G. et al. Accelerated topological design of metaporous materials of broadband sound absorption performance by generative adversarial networks. Materials & Design, 2021, 207: 109855

[158] Chen C H, Chen K Y, Shu Y C. Data-driven bio-mimetic composite design: direct prediction of stress-strain curves from structures using cGANs. Journal of the Mechanics and Physics of Solids, 2024, 193: 105857

[159] Liu Y J, He H L, Cao Y J. et al. Inverse design of TPMS piezoelectric metamaterial based on deep learning. Mechanics of Materials, 2024, 198: 105109

[160] Zhang Z W, Zhang Y Y, Wen Y T. et al. Intelligent defect detection method for additive manufactured lattice structures based on a modified YOLOv3 model. Journal of Nondestructive Evaluation, 2022, 41(1): 3

[161] Xue T J, Wallin T J, Menguc Y. et al. Machine learning generative models for automatic design of multi-material 3D printed composite solids. Extreme Mechanics Letters, 2020, 41: 100992

[162] Yasuda H, Yamaguchi K, Miyazawa Y. et al. Data-driven prediction and analysis of chaotic origami dynamics. Communications Physics, 2020, 3(1): 168

[163] Wang L W, Chan Y C, Ahmed F. et al. Deep generative modeling for mechanistic-based learning and design of metamaterial systems. Computer Methods in Applied Mechanics and Engineering, 2020, 372: 113377

[164] Feng R Q, Liu F C, Xu W J. et al. Topology optimization method of lattice structures based on a genetic algorithm. International Journal of Steel Structures, 2016, 16(3): 743–753

[165] Park C, Lee S. Tunable anisotropy in lattice structures via deep learning-based optimization. International Journal of Mechanical Sciences, 2025, 290: 110121

[166] Kuszczak I, Azam F I, Bessa M A. et al. Bayesian optimisation of hexagonal honeycomb metamaterial. Extreme Mechanics Letters, 2023, 64: 102078

[167] Gao Z, Shi Y, Ma L, Du J, Qiu J. Bioinspired hierarchical composite for multiband defense. ACS Applied Materials and Interfaces, 2024, 16(37): 49687–49700

[168] Maurizi M, Gao C, Berto F. Predicting stress, strain and deformation fields in materials and structures with graph neural networks. Scientific Reports, 2022, 12(1): 21834

[169] Vafaeefar M, Moerman K M, Vaughan T J. LatticeWorks: an open-source MATLAB toolbox for nonuniform, gradient and multi-morphology lattice generation, and analysis. Materials & Design, 2025, 250: 113564

[170] Al-Ketan O, Abu Al-Rub R K. MSLattice: a free software for generating uniform and graded lattices based on triply periodic minimal surfaces. Material Design & Processing Communications, 2021, 3(6): e205

[171] Chris-Amadin H, Ibhadode O. LattGen: a TPMS lattice generation tool. Software Impacts, 2024, 21: 100665

[172] Marten. STL Lattice Generator. MATLAB Central File Exchange, version 1.4.0.0, 27 Jan 2015. Available online (accessed 8 Jul 2026)

[173] Gleadall A. FullControl GCode Designer: open-source software for unconstrained design in additive manufacturing. Additive Manufacturing, 2021, 46: 102109

[174] Carbon, Inc. Carbon Design Engine. Available online (accessed 8 Jul 2026)

[175] McGill Additive Design and Manufacturing Laboratory. Intralattice: open-source lattice structure design software for AM. Available online (accessed 8 Jul 2026)

[176] nTop. Lattice structures. Available online (accessed 8 Jul 2026)

[177] Materialise. Materialise Magics: 3D printing data and build preparation software. Available online (accessed 8 Jul 2026)

[178] Altair Engineering Inc. Lattice Structure Optimization. OptiStruct User Guide. Available online (accessed 8 Jul 2026)

[179] Autodesk, Inc. Autodesk Within Medical Technical Manual. Autodesk, 2016. Available online (accessed 8 Jul 2026)

[180] ParaMatters, Inc. ParaMatters 4.0 generative design software opens the door for digital and traditional manufacturers to automate design-to-manufacturing processes. Business Wire, 2020. Available online (accessed 8 Jul 2026)

[181] Monolith. Monolith: voxel-based modeling engine for multimaterial 3D printing. Available online (accessed 8 Jul 2026)

[182] Li Z, Xiao W, Xiong C, Wang S. gpLattice: from generation to optimization, an efficient and flexible lattice generation software kernal. In: Proceedings of CAD'23. Mexico City, Mexico, 2023: 281-285

[183] Forward AM. Ultrasim 3D Lattice Library. Available online (accessed 8 Jul 2026)

[184] LatticeRobot. LatticeRobot. Available online (accessed 8 Jul 2026)

[185] Jadhav Y, Berthel J, Hu C S. et al. Generative lattice units with 3D diffusion for inverse design: GLU3D. Advanced Functional Materials, 2024, 34(41): 2404165

[186] Wang C, Tan X P, Tor S B. et al. Machine learning in additive manufacturing: state-of-the-art and perspectives. Additive Manufacturing, 2020, 36: 101538

[187] Salari-Sharif L, Godfrey S W, Tootkaboni M. et al. The effect of manufacturing defects on compressive strength of ultralight hollow microlattices: a data-driven study. Additive Manufacturing, 2018, 19: 51–61

[188] Zhang Z Y, Fang H, Xu Z. et al. Multi-objective generative design of three-dimensional material structures. APL Machine Learning, 2023, 1(4): 046120

[189] Dong H W, Shen C, Liu Z. et al. Inverse design of phononic meta-structured materials. Materials Today, 2024, 80: 824–855

[190] Chen J J, Duan H L, Huang G L. Transformer-based inverse-design model for optimal multilayer microperforated panels. Physical Review Applied, 2025, 23(2): 024044

[191] Chen J Z, Wu H, Zhou J Z. et al. Heterostructured mechanical metamaterials inspired by the shell of Strombus gigas. Journal of the Mechanics and Physics of Solids, 2024, 188: 105658

[192] Mao A R, Zhao N F, Liang Y H. et al. Mechanically efficient cellular materials inspired by cuttlebone. Advanced Materials, 2021, 33(15): 2007348

[193] Cheng X, Fan Z C, Yao S L. et al. Programming 3D curved mesosurfaces using microlattice designs. Science, 2023, 379(6638): 1225–1232

[194] Deng W T, Kumar S, Vallone A. et al. AI-enabled materials design of non-periodic 3D architectures with predictable direction-dependent elastic properties. Advanced Materials, 2024, 36(34): 2308149

[195] Wang H Y, Yang Y S, Zhou X Q. et al. Rational design of mechanical bio-metamaterials for biomedical applications. Progress in Materials Science, 2026, 156: 101545

[196] Roh Y, Lee S, Won S M. et al. Crumple-recoverable electronics based on plastic to elastic deformation transitions. Nature Electronics, 2024, 7(1): 66–76

[197] Dong Z C, Ren X Y, Jia B. et al. Composite patch with negative Poisson’s ratio mimicking cardiac mechanical properties: design, experiment and simulation. Materials Today Bio, 2024, 26: 101098

[198] Meng X S, Zhou L C, Liu L. et al. Deformable hard tissue with high fatigue resistance in the hinge of bivalve Cristaria plicata. Science, 2023, 380(6651): 1252–1257

[199] Cheng X, Shen Z M, Zhang Y H. Bioinspired 3D flexible devices and functional systems. National Science Review, 2024, 11(3): nwad314

[200] Lu J F, Deng J J, Wei Y. et al. Hierarchically mimicking outer tooth enamel for restorative mechanical compatibility. Nature Communications, 2024, 15(1): 10182

[201] Singh A, Al-Ketan O, Karathanasopoulos N. Hybrid manufacturing and mechanical properties of architected interpenetrating phase metal-ceramic and metal-metal composites. Materials Science and Engineering: A, 2024, 897: 146322

[202] Zhang M Y, Zhao N, Yu Q. et al. On the damage tolerance of 3-D printed Mg-Ti interpenetrating-phase composites with bioinspired architectures. Nature Communications, 2022, 13(1): 3247

[203] Tankasala H C, Fleck N A. The crack growth resistance of an elastoplastic lattice. International Journal of Solids and Structures, 2020, 188–189: 233–243

[204] Song Z Q, Wu K J, Wang Z W. et al. Fracture mechanics of bi-material lattice metamaterials. Journal of the Mechanics and Physics of Solids, 2024, 192: 105835

[205] Maurizi M, Edwards B W, Gao C. et al. Fracture resistance of 3D nano-architected lattice materials. Extreme Mechanics Letters, 2022, 56: 101883

[206] Shaikeea A J D, Cui H C, O’Masta M. et al. The toughness of mechanical metamaterials. Nature Materials, 2022, 21(3): 297–304

[207] Surjadi J U, Lu Y. Design criteria for tough metamaterials. Nature Materials, 2022, 21(3): 272–274

[208] Zhang W H, Tang C H. Lightweighting of aerospace and aeronautical equipment: challenges and perspectives. Acta Aeronautica et Astronautica Sinica, 2024, 45(5): 529965

[209] Lu T J, He D P, Chen C Q. et al. The multi-functionality of ultra-light porous metals and their applications. Advances in Mechanics, 2006, 36(4): 517–535

[210] Oh D K, Lee T, Ko B. et al. Nanoimprint lithography for high-throughput fabrication of metasurfaces. Frontiers of Optoelectronics, 2021, 14(2): 229–251

[211] Yoon G, Kim I, Rho J. Challenges in fabrication towards realization of practical metamaterials. Microelectronic Engineering, 2016, 163: 7–20

[212] Ouyang W Q, Xu X Y, Lu W P. et al. Ultrafast 3D nanofabrication via digital holography. Nature Communications, 2023, 14(1): 1716

[213] Zhao Z L, Liu Y Y, Wang P. Computational design of bio-inspired mechanical metamaterials based on lipidic cubic phases. JOM, 2023, 75(7): 2126–2136

[214] Yi S Z, Wang L, Chen Z P. et al. High-throughput fabrication of soft magneto-origami machines. Nature Communications, 2022, 13(1): 4177

[215] Surjadi J U, Portela C M. Enabling three-dimensional architected materials across length scales and timescales. Nature Materials, 2025, 24(4): 493–505

[216] Burley S K. An overview of structural genomics. Nature Structural Biology, 2000, 7(11): 932–934

[217] Yu W B. Structure genome: fill the gap between materials genome and structural analysis. In: 56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference. Kissimmee, Florida, 2015, AIAA 2015-0201

[218] McClung A, Torfeh M, Einck V J. et al. Visible metalenses with high focusing efficiency fabricated using nanoimprint lithography. Advanced Optical Materials, 2024, 12(9): 2301865

[219] Baron A, Aradian A, Ponsinet V. et al. Bottom-up nanocolloidal metamaterials and metasurfaces at optical frequencies. Comptes Rendus Physique, 2020, 21(4-5): 443–465

[220] Geng Q, Wang D, Chen P. et al. Ultrafast multi-focus 3-D nano-fabrication based on two-photon polymerization. Nature Communications, 2019, 10: 2179

[221] Jin H, Espinosa H D. Mechanical metamaterials fabricated from self-assembly: a perspective. Journal of Applied Mechanics, 2024, 91(4): 040801

[222] Xu B, Lin X, Mei Y. Versatile rolling origami to fabricate functional and smart materials. Cell Reports Physical Science, 2020, 1(11): 100244

[223] Jin H, Zhang B, Cao Q. et al. Characterization and inverse design of stochastic mechanical metamaterials using neural operators. Advanced Materials, 2025, 37(29): 2420063

[224] Zhang C D, Wang B, Zhu H Y. et al. Structure genome based machine learning method for woven lattice structures. International Journal of Mechanical Sciences, 2023, 245: 108134

[225] Kalidindi S R, Niezgoda S R, Landi G, et al. A novel framework for building materials knowledge systems. Computers, Materials & Continua, 2010, 17(2): 103–126

[226] Lookman T, Balachandran P V, Xue D Z. et al. Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design. npj Computational Materials, 2019, 5(1): 21

[227] Gopakumar A M, Balachandran P V, Xue D Z. et al. Multi-objective optimization for materials discovery via adaptive design. Scientific Reports, 2018, 8(1): 3738

[228] MacLeod B P, Parlane F G L, Morrissey T D. et al. Self-driving laboratory for accelerated discovery of thin-film materials. Science Advances, 2020, 6(20): eaaz8867

[229] Gomes C P, Bai J W, Xue Y X. et al. CRYSTAL: a multi-agent AI system for automated mapping of materials’ crystal structures. MRS Communications, 2019, 9(2): 600–608

[230] Lin D Z, Pan K J, Li Y Y. et al. A high-throughput experimentation platform for data-driven discovery in electrochemistry. Science Advances, 2025, 11(14): eadu4391

[231] Dave A, Mitchell J, Kandasamy K. et al. Autonomous discovery of battery electrolytes with robotic experimentation and machine learning. Cell Reports Physical Science, 2020, 1(12): 100264

[232] Mannix A J, Ye A, Sung S H. et al. Robotic four-dimensional pixel assembly of van der Waals solids. Nature Nanotechnology, 2022, 17(4): 361–366

[233] Cai Y H, Xiong J, Chen H. et al. A review of in-situ monitoring and process control system in metal-based laser additive manufacturing. Journal of Manufacturing Systems, 2023, 70: 309–326

[234] Butler K T, Davies D W, Cartwright H. et al. Machine learning for molecular and materials science. Nature, 2018, 559(7715): 547–555

[235] Häse F, Roch L M, Kreisbeck C. et al. Phoenics: a Bayesian optimizer for chemistry. ACS Central Science, 2018, 4(9): 1134–1145

[236] Callaham J L, Koch J V, Brunton B W. et al. Learning dominant physical processes with data-driven balance models. Nature Communications, 2021, 12(1): 1016

[237] Jiang Y F, Cao S Y, Meng H. et al. A data-driven design for sound absorption of acoustic metamaterials based on large language models. Scientific Reports, 2025, 16(1): 517

[238] Himanen L, Geurts A, Foster A S. et al. Data-driven materials science: Status, challenges, and perspectives. Advanced Science, 2019, 6(21): 1900808

[239] Wang L, Ma C, Feng X Y. et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 2024, 18(6): 186345

[240] Yin S K, Fu C Y, Zhao S R. et al. A survey on multimodal large language models. National Science Review, 2024, 11(12): nwae403

[241] Li X, Wang S, Zeng S. et al. A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges. Vicinagearth, 2024, 1(1): 9

[242] Yang C, Zhu Y, Lu W. et al. Survey on knowledge distillation for large language models: methods, evaluation, and application. ACM Transactions on Intelligent Systems and Technology, 2025, 16(6): 143

[243] OpenAI. Using GPT-5.5. OpenAI API documentation. Available online (accessed 23 Jun 2026)

[244] Anthropic. Models overview. Claude API documentation. Available online (accessed 23 Jun 2026)

[245] Google DeepMind. Gemini 3.5: frontier intelligence with action. Available online (accessed 23 Jun 2026)

[246] DeepSeek. DeepSeek V4 preview release. DeepSeek API documentation. Available online (accessed 23 Jun 2026)

[247] Yang A, Li A, Yang B, et al. Qwen3 Technical Report. arXiv preprint, 2025, arXiv: 2505.09388

[248] Meta AI. The Llama 4 herd: the beginning of a new era of natively multimodal AI innovation. Meta AI Blog. Available online (accessed 23 Jun 2026)

[249] Mistral AI. Models overview. Mistral AI documentation. Available online (accessed 23 Jun 2026)

[250] Brodnik N R, Carton S, Muir C. et al. Perspective: large language models in applied mechanics. Journal of Applied Mechanics, 2023, 90(10): 101008

[251] Mustapha K B. A survey of emerging applications of large language models for problems in mechanics, product design, and manufacturing. Advanced Engineering Informatics, 2025, 64: 103066

[252] Tian J, Sobczak M T, Patil D, et al. A multi-agent framework integrating large language models and generative AI for accelerated metamaterial design. arXiv preprint: arXiv: 2503.19889, 2025

[253] Ni B, Buehler M J. Mechagentss: large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. Extreme Mechanics Letters, 2024, 67: 102131

[254] Lu J X, Li H R, Ning F W, et al. Constructing mechanical design agent based on large language models. arXiv preprint: arXiv: 2408.02087, 2024

[255] Jadhav Y, Farimani A B. Large language model agent as a mechanical designer. Journal of Engineering Design, 2025. doi: 10.1080/09544828.2026.2624356

[256] Chong Y Y, Feng S, Wang S. et al. Large model-driven, human-computer collaborative robotic AI-chemist cloud facility. Bulletin of Chinese Academy of Sciences, 2024, 39(1): 41–49

[257] Ni Z Q, Li Y H, Hu K J, et al. MatPilot: an LLM-enabled AI materials scientist under the framework of human-machine collaboration. arXiv preprint: arXiv: 2411.08063, 2024

[258] Hu S G, Li M Y, Xu J W, et al. Electromagnetic metamaterial agent. Light: Science & Applications, 2025, 14(1): 12

[259] Feynman R P. There’s plenty of room at the bottom. Engineering and Science, 1960, 23(5): 22–36

[260] Roco M C, Mirkin C A, Hersam M C. Nanotechnology research directions for societal needs in 2020: retrospective and outlook. Dordrecht: Springer, 2011

[261] Jiang Z L. Advance from biomechanics to mechanobiology. Advances in Mechanics, 2017, 47(1): 309–332

[262] Fung Y C. Biomechanics: mechanical properties of living tissues. 2nd ed. New York: Springer, 1993

[263] Humphrey J D. Review Paper: continuum biomechanics of soft biological tissues. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2003, 459(2029): 3–46

[264] Xue C, Wang Y. Progress in corneal mechanobiology research. Chinese Journal of Optometry Ophthalmology and Visual Science, 2022, 24(7): 556–560

[265] Clyne A M, Marcolongo M, Darling E M. et al. Translating mechanobiology to the clinic: a panel discussion from the 2018 CMBE Conference. Cellular and Molecular Bioengineering, 2018, 11(6): 531–535

[266] Gefen A, Weihs D. Cytoskeleton and plasma-membrane damage resulting from exposure to sustained deformations: a review of the mechanobiology of chronic wounds. Medical Engineering & Physics, 2016, 38(9): 828–833

[267] Goonoo N. Tunable biomaterials for myocardial tissue regeneration: promising new strategies for advanced biointerface control and improved therapeutic outcomes. Biomaterials Science, 2022, 10(7): 1626–1646

[268] Zou G J, Sow C H, Wang Z S. et al. Mechanomaterials and nanomechanics: toward proactive design of material properties and functionalities. ACS Nano, 2024, 18(18): 11492–11502

[269] Cai P Q, Wang C X, Gao H J. et al. Mechanomaterials: a rational deployment of forces and geometries in programming functional materials. Advanced Materials, 2021, 33(46): 2007977

[270] Gotla S, Tong C, Matysiak S. Load-bearing nanostructures in composites of chitosan with anionic surfactants: implications for programmable mechanomaterials. ACS Applied Nano Materials, 2022, 5(5): 6463–6473

[271] Wu W W, Xia R, Qian G A. et al. Mechanostructures: rational mechanical design, fabrication, performance evaluation, and industrial application of advanced structures. Progress in Materials Science, 2023, 131: 101021

[272] Lu T J. What is MechanoEngineering? MechanoEngineering, 2026, 1(1): 010401

[273] Wang Y J, Zou G J, Gao H J. Mechano-X: a paradigm for mechanics-based interdisciplinary innovation. MechanoEngineering, 2026, 1(1): 010801

RIGHTS & PERMISSIONS

The Author(s). This article is published by Higher Education Press.