From Crystalline Frameworks to Dynamic Networks: Artificial Intelligence-Guided Design of Metal–Organic Materials

Abstract

Artificial intelligence has greatly accelerated the design and screening of metal–organic materials, particularly for crystalline systems with well-defined topologies and increasingly standardized structural databases. However, this success has also created a structure-centric design paradigm that is less suitable for metal–organic systems whose functions are governed by process history, interfacial assembly, and dynamic coordination rather than by a single idealized lattice. This Perspective proposes that artificial intelligence (AI)-guided design of metal–organic materials should expand beyond crystalline metal–organic frameworks (MOFs) to encompass a broader structural continuum, ranging from long-range ordered frameworks to dynamic, non-periodic coordination networks. Metal–polyphenol networks (MPNs) are used here as an experimentally tractable example within a broader family of structurally dynamic metal–organic materials, as they arise from coordination interactions between metal ions and polyphenolic ligands, generally lack long-range crystallographic periodicity, and exhibit functions that are governed by interfacial assembly, environmental responsiveness, and pathway-dependent structural evolution. These features challenge conventional descriptor design and database-driven prediction, but also create opportunities for AI approaches that are process-aware, interface-sensitive, and function-oriented. By placing MOFs and MPNs within a unified framework of structural order, this Perspective outlines how machine learning, multimodal characterization, active learning, and closed-loop experimentation could expand metal–organic materials design from topology prediction toward dynamic network optimization.

1. Introduction

Metal–organic materials have become an important class of functional materials because their structures and properties can be regulated through the coordination between metal centers and organic ligands [1]. Among them, metal–organic frameworks (MOFs) have attracted particular attention due to their crystalline structures, well-defined topologies, tunable pores, and broad applications in adsorption, separation, catalysis, sensing, energy conversion, and biomedicine [2,3,4]. The long-range order of MOFs has also made them highly compatible with data-driven materials design. Crystallographic information files, pore descriptors, topological information, linker chemistry, metal-node identity, and calculated properties have provided a structured data foundation for machine learning and high-throughput screening [5].

Artificial intelligence (AI) has therefore greatly accelerated the design and screening of MOFs. Large databases and computational workflows have enabled the rapid evaluation of adsorption capacity, selectivity, stability, catalytic activity, and other target properties. In this context, MOFs represent one of the most successful examples of AI-assisted design in metal–organic materials. However, this success has also shaped a relatively narrow view of what metal–organic materials are in the context of AI. Most current models are optimized for systems that can be represented as static, crystalline, and topologically well-defined frameworks [6].

This view has been highly productive, but it also points to the need for a broader framework that includes less ordered and more dynamic metal–organic systems [7]. Metal–organic materials are not limited to crystalline MOFs [8,9]. Many functional systems are assembled through dynamic coordination, ligand exchange, interfacial binding, solvent-dependent restructuring, or pathway-dependent growth [10]. These materials may be amorphous, short-range ordered, mesoscopically organized, or non-periodic. Their functions are often governed not by a fixed topology, but by local coordination environments, assembly history, interfacial structure, and environmental responsiveness. In such systems, the relevant design variables extend beyond metal nodes and organic linkers to include assembly history, interfacial context, and operating conditions.

This Perspective places crystalline MOFs and dynamic metal–polyphenol networks (MPNs) within a unified structural continuum of metal–organic materials. MOFs represent long-range ordered crystalline frameworks, while MPNs represent dynamic metal–organic networks formed through coordination between metal ions and polyphenolic ligands. MPNs are positioned not simply as disordered coordination materials, but as one representative and experimentally accessible testbed for assessing how AI-guided materials design can move beyond static structure prediction toward the optimization of process-dependent, dynamically evolving material states. The comparison between these two systems is not intended to replace the established MOF design paradigm. Instead, it is used to clarify how AI-guided design may expand from topology-centered prediction to process-aware and function-oriented optimization. Accordingly, this Perspective presents an integrated design logic for representing metal–organic materials through their composition, assembly pathway, operating environment, and intended functional state.

2. Structural Order as a Design Axis for Metal–Organic Materials

Metal–organic materials are often introduced through composition, namely the combination of metal centers and organic ligands [11]. For AI-guided design, however, composition alone is not sufficient. Two materials built from similar coordination motifs may require fundamentally different data representations if one forms a long-range crystalline framework and the other forms a hydrated, interfacial, and dynamically rearranging network. Structural order therefore provides a useful axis along which metal–organic materials can be discussed, described, and ultimately designed [12]. As illustrated in Figure 1, this structural continuum spans long-range ordered crystalline frameworks, partially ordered intermediate systems, and dynamic non-periodic coordination networks.

At one end of this axis are crystalline metal–organic frameworks. Their structures can usually be expressed by unit-cell parameters, topology, pore metrics, metal-node connectivity, and organic linker identity. Such representations are compatible with existing crystallographic databases and with the descriptor systems commonly used in machine learning [13]. At the other end are dynamic metal–organic networks, including metal–polyphenol networks, in which long-range crystallographic periodicity is generally absent. These systems are not structureless. Rather, their order is expressed locally through metal–ligand coordination, interfacial organization, short-range packing, hydration, and assembly history [14,15]. Between these two limits lies a broad intermediate region. Defective MOFs, amorphous MOFs, low-dimensional coordination polymers, mesocrystalline assemblies, and partially ordered hybrid networks all occupy positions between ideal crystallinity and dynamic non-periodicity [16]. These materials demonstrate that structural order is not a binary concept. A material may lose long-range periodicity while retaining chemically meaningful local coordination. It may show mesoscale organization without being crystallographically ordered. It may also undergo structural rearrangement during synthesis, activation, or use.

2.1. Crystallinity Is Not the Only Form of Structural Information

The dominance of MOF databases has understandably encouraged a crystallographic view of metal–organic materials. This view is powerful for crystalline frameworks, but it becomes incomplete when applied to dynamic networks. In MPNs, for example, the most important structural information may not be a periodic lattice, but the distribution of coordination environments, the density of metal–phenolic crosslinks, the degree of ligand oxidation, the interfacial coverage, and the response of the network to pH or ionic strength [17]. For AI models, this distinction is critical. Crystalline MOFs can often be encoded through graph representations derived from idealized crystal structures, utilizing approaches such as crystal graph convolutional neural networks [18]. MPNs require descriptors that capture coordination chemistry and process history. Useful descriptors may include the identity and valence state of the metal ion, the number and arrangement of phenolic groups, ligand denticity, metal-to-ligand ratio, pH, solvent composition, assembly time, substrate surface chemistry, and functional readouts such as coating stability, permeability, catalytic activity, or biological protection.

2.2. A Continuum Rather than a Hierarchy

It is tempting to arrange metal–organic materials in a hierarchy from “ordered” to “disordered”, with crystalline MOFs viewed as ideal and non-periodic networks as less defined. Such a hierarchy is misleading. Structural order is not necessarily equivalent to functional superiority. In some cases, long-range order is essential for selective adsorption, molecular sieving, or crystallographic modeling. In other cases, dynamic coordination and structural adaptability are precisely what enable function [19]. MPNs illustrate this point clearly. Their non-periodic and adaptive character allows them to form conformal coatings on particles, proteins, cells, hydrogels, membranes, and irregular biological interfaces [20,21]. Their functions often arise from the ability to reorganize, bind, protect, release, or respond under changing environmental conditions. From this perspective, structural disorder is not simply an obstacle for AI. It is a design variable that needs to be described at the appropriate length scale.

Beyond MPNs, similar representation challenges arise in a range of dynamic or partially ordered metal–organic systems. Amorphous MOFs and MOF glasses may lose long-range periodicity while retaining local metal–ligand connectivity and chemically meaningful short-range order. Defective MOFs and quasi-MOFs can exhibit process-dependent defect distributions, missing-linker chemistry, and activation-sensitive structures. Low-dimensional coordination polymers, coordination gels, and supramolecular metal–ligand networks often depend on solvent, concentration, counterions, ligand exchange, and assembly sequence. Biohybrid MOF composites, adaptive metal–organic coatings, and hydrogel–MOF systems further illustrate how biological interfaces, hydration, mechanical softness, and operating environments can become part of the functional material state. These examples indicate that MPNs should be viewed not as the sole model of dynamic metal–organic networks, but as an experimentally accessible testbed within a broader class of process-dependent and non-periodic coordination materials.

2.3. Implications for AI Representation

The structural order continuum suggests that AI-guided design should not rely on a single universal representation. For highly crystalline MOFs, topology, pore structure, linker chemistry, and metal-node identity remain central. For defective or amorphous MOFs, these descriptors should be supplemented with defect density, missing-linker chemistry, activation history, local spectroscopy, and stability data, enabling the development of machine learning potentials capable of describing spatial disorder up to the mesoscale [22]. For MPNs and other dynamic networks, process-encoded descriptors and multimodal characterization become indispensable. A more general AI framework for metal–organic materials should therefore connect four types of information: composition, structure, process, and function. Crystalline MOFs are currently rich in structural information, but often weaker in process and failure data, necessitating multimodal approaches that directly connect synthesis precursors and characterization to applications [23]. MPNs are frequently rich in functional observations, but weaker in standardized structural data. Bridging these asymmetries is a major opportunity for the next stage of AI-guided metal–organic materials design.

3. Crystalline MOFs and the Success of Topology-Centered AI

MOFs have emerged as a model platform for AI-assisted materials design because they combine vast chemical diversity with a high degree of structural definition. Their modular construction from metal nodes or clusters and organic linkers allows nearly unlimited combinations of compositions, connectivities, pore environments, and functional groups. At the same time, their crystalline nature enables these structural features to be represented in relatively standardized and machine-readable forms, such as crystallographic information files, topological nets, pore descriptors, metal-node identities, and linker-based molecular descriptors. The success of MOFs in AI is grounded in three enabling conditions that extend beyond their broad materials significance, namely machine-readable crystallographic structures, scalable databases, and relatively stable structure–property relationships. Together with their compositional tunability and structural regularity, these features distinguish MOFs from many other porous or coordination materials. This allows large experimental and hypothetical MOF libraries to be constructed, annotated, and screened computationally [24]. Consequently, MOFs are particularly well suited for machine learning models, high-throughput virtual screening, graph-based representations, generative design, and inverse design strategies aimed at identifying materials with target adsorption, separation, catalytic, electronic, or stability properties, thereby shifting the paradigm from trial-and-error discovery to property-directed synthesis [25]. These advantages make MOFs the starting point for AI-guided metal–organic materials design, but they also define the boundary conditions under which topology-centered AI is most reliable.

3.1. Why MOFs Are Well Suited to Data-Driven Design

A crystalline MOF can often be reduced to a set of machine-readable structural and chemical descriptors. These include metal identity, linker structure, coordination connectivity, topology, pore limiting diameter, largest cavity diameter, accessible surface area, void fraction, density, and functional groups. Many of these descriptors can be extracted automatically from crystallographic information files or computed through established workflows. As a result, MOFs provide a data environment in which AI models can be trained at scale. This data foundation has enabled rapid progress in property prediction. Machine learning models have been used to predict gas adsorption capacity, separation selectivity, mechanical stability, solvent stability, catalytic activity, electronic properties, and thermal behavior [26]. In parallel, generative models and inverse design strategies have begun to propose hypothetical frameworks with target pore structures or functional properties, utilizing advanced architectures like deep artificial neural networks to navigate immense compositional spaces [27]. These developments demonstrate the strength of topology-centered AI when structural information is abundant and standardized.

3.2. The Limits of Idealized Crystalline Inputs

The success of AI in MOF research should not obscure the assumptions on which many models depend. In most workflows, the input structure is treated as a static and idealized crystal. Defects, disorder, particle size, phase purity, solvent residues, activation history, surface reconstruction, and environmental degradation are often simplified or excluded [28]. In many screening studies, synthesizability and long-term stability are considered only after computational ranking. This limitation is not unique to MOFs, but it is especially important for their transition from virtual candidates to working materials. A framework predicted to have excellent adsorption or catalytic performance may be difficult to synthesize, unstable during activation, sensitive to moisture, or altered by defects. Conversely, a material with moderate idealized performance may become attractive because it is robust, scalable, and tolerant to real operating conditions [29].

3.3. From Structure Prediction to Materials Realization

For crystalline MOFs, the next challenge is not only to predict properties from structures, but also to predict whether those structures can be made, activated, maintained, and used. This requires the integration of synthesis recipes, solvent conditions, modulator effects, temperature, reaction time, activation procedures, and post-synthetic modifications into AI models. Natural language processing can help extract synthesis information from the literature [30], while robotic and self-driving platforms can generate more systematic experimental data [31]. In this sense, even the MOF field is moving beyond purely topology-centered prediction. This transition is central to the logic of the present Perspective, as the need for synthesis-aware and stability-aware AI in crystalline MOFs further underscores the necessity of a broader framework for dynamic networks such as MPNs, where process, interface, and function must be treated as primary descriptors rather than secondary metadata. MOFs provide the mature starting point; MPNs expose the limits of a static structure-only paradigm.

4. Dynamic MPNs and the Challenge of Process-Dependent Design

MPNs occupy a distinct region of the metal–organic materials landscape, where function is governed less by long-range crystallographic periodicity than by dynamic coordination, interfacial assembly, and environment-dependent structural evolution. They are especially useful for extending AI-guided design because they combine a chemically interpretable coordination basis with a highly process-sensitive and interface-dependent material state. They are typically constructed under mild aqueous or mixed-solvent conditions through coordination interactions between metal ions and polyphenolic ligands. Common polyphenolic building blocks include tannic acid, gallic acid, catechol-containing molecules, epigallocatechin gallate, ellagic acid, and related phenolic structures, whose catechol, galloyl, pyrogallol, or phenolate groups provide multiple coordination sites. The metal components can range from biologically relevant ions such as Fe³⁺, Zn²⁺, Ca²⁺, Mg²⁺, and Mn²⁺ to redox-active or catalytically functional centers, each contributing different coordination geometries, binding strengths, hydrolysis behaviors, redox properties, and biological effects. As a result, MPN formation is strongly influenced by pH, metal-to-ligand ratio, ligand oxidation state, ionic strength, solvent environment, assembly time, and substrate chemistry [32]. Rather than producing a fixed periodic framework, these variables usually give rise to hydrated, non-stoichiometric, interfacial, and dynamically reconfigurable coordination networks. Such structural features make MPNs particularly suitable for conformal coatings, biointerfaces, encapsulation of delicate biological entities, responsive delivery, catalytic interfaces, and protective layers [33], but they also make them difficult to describe using conventional crystallographic descriptors developed for MOFs.

4.1. MPNs Are Not Simply Disordered MOFs

MPNs should not be described merely as disordered analogs of MOFs. Their formation mechanism, structural expression, and functional logic are different. A MOF is usually designed through reticular chemistry to generate a periodic framework. An MPN is commonly assembled through dynamic metal–phenolic coordination at an interface, in solution, or around a template [34]. The resulting material may not exhibit long-range crystallographic order, but it can still possess well-defined local coordination environments and reproducible functional behavior. This distinction matters for AI. A model trained to optimize MOF topology may search for pore geometry, linker length, or metal-node connectivity. A model for MPNs must instead consider assembly chemistry, substrate interaction, network growth, local coordination, hydration, and environmental response. The design target also changes. In MPNs, the desired output may be coating integrity, protein protection, cell surface functionalization, catalytic selectivity, oxidative stability, controlled release, or biological compatibility rather than gas uptake or framework porosity [35].

4.2. Pathway-Dependent Function and the Descriptor Problem

The function of MPNs is often pathway-dependent. Small changes in pH, metal-to-ligand ratio, solvent composition, ionic strength, oxidation state, or assembly sequence can alter coating thickness, network density, permeability, stability, and release behavior. The same metal and polyphenol can therefore produce different functional outcomes under different assembly conditions. This feature makes MPNs challenging for conventional database-driven prediction. There is usually no single “final structure” that fully determines function. Instead, function emerges from a combination of local coordination chemistry and assembly pathway. For AI-guided design, the input should therefore include how the material was made, where it was assembled, and under what conditions it is expected to function [36]. This pathway dependence means the key bottleneck for AI-guided MPN design is not simply the lack of data. It is the lack of standardized descriptors that connect chemistry, process, structure, and function. For example, the term “Fe–TA network” may refer to different materials depending on pH, Fe³⁺/TA ratio, assembly time, substrate, buffer, oxidation state, and washing conditions [17]. Without recording these variables, the same nominal composition can represent substantially different networks [37]. Useful MPN descriptors should therefore include metal ion type, valence state, coordination preference, ligand structure, phenolic group density, ligand oxidation susceptibility, binding strength, metal-to-ligand ratio, pH, ionic strength, solvent, assembly time, temperature, substrate chemistry, coating thickness, surface charge, and stability under relevant environments. Functional descriptors should be selected according to the intended application, such as protection efficiency, adhesion strength, permeability, catalytic turnover, antioxidant capacity, antimicrobial activity, or release kinetics.

4.3. Why MPNs Are a Timely Test Case for AI

MPNs are particularly suitable for expanding AI-guided metal–organic materials design because they combine chemical simplicity with process complexity. Their building blocks are often accessible, biocompatible, and chemically diverse. Their assembly can be rapid and modular. At the same time, their structure cannot be fully described by crystallography, and their performance is sensitive to processing and interfaces. This combination makes MPNs a stringent but experimentally tractable test case for small-data, chemistry-aware, process-encoded, and function-oriented AI methods [38]. Representative examples of pH-responsive coordination, compositional and morphological diversity, pathway-dependent evolution, interfacial assembly, and AI-enabled optimization are summarized in Figure 2. Progress in this area would advance the design of MPNs and, more broadly, establish transferable principles for amorphous coordination polymers, biointerfacial metal–organic assemblies, adaptive hybrid coatings, and other metal–organic materials whose performance is governed by dynamic rather than periodic structures.

The generality of the MPN case should therefore be understood at the level of representation and workflow rather than at the level of identical chemistry. The broadly transferable elements include the need to encode local coordination environments, assembly history, interfacial context, environmental response, multimodal characterization, and application-specific function. These concepts are expected to be relevant to amorphous coordination polymers, coordination gels, supramolecular metal–ligand networks, biohybrid metal–organic assemblies, adaptive coatings, and other process-dependent metal–organic systems. By contrast, several descriptors remain especially grounded in MPN chemistry, including phenolic ligand structure, metal–phenolic binding strength, ligand oxidation state, pH-dependent coordination equilibria, and conformal film growth on soft or biological interfaces. Distinguishing these transferable and MPN-specific features is important for adapting the proposed AI framework to other dynamic metal–organic materials without overgeneralizing from a single chemical platform.

5. From Topology Prediction to Dynamic Network Optimization

The extension of AI-guided design from crystalline MOFs to dynamic metal–organic networks requires more than adding new material examples to existing datasets. It calls for a change in how metal–organic materials are described, how design variables are selected, and how functional performance is connected to structure and process. In crystalline MOFs, topology, pore geometry, linker structure, and metal-node identity often provide a powerful basis for property prediction [39]. In dynamic networks, such as MPNs and related amorphous, supramolecular, or biohybrid metal–organic systems, these descriptors must be complemented by information on local coordination, assembly conditions, interfacial interactions, and environmental responses [40]. Here, dynamic network optimization is defined as an AI-guided design strategy that focuses on process-dependent material states rather than single idealized crystal structures. These states are determined by composition, local coordination, assembly pathway, interface, operating environment, and functional output, and are optimized through four interconnected modules, including process-encoded material identity, interface-sensitive descriptors, multimodal small-data learning, and closed-loop experimental optimization.

Figure 2. Structural dynamics, interfacial assembly, and data-driven optimization of MPNs. (A) pH-dependent dynamic coordination. Visual color changes and corresponding UV-vis spectra illustrate the transition of local metal–ligand coordination states, demonstrating that MPN structures are highly sensitive to processing conditions rather than adopting fixed periodic lattices [17]. (B) Compositional and morphological versatility. Electron microscopy and energy-dispersive X-ray spectroscopy (EDS) elemental mapping of MPN capsules assembled from various metal ions (e.g., Cu, Al, Zr) [8]. (C,D) Spatiotemporal monitoring of MPNs, showcasing the pathway-dependent nature of dynamic network assembly and structural evolution [41]. (E) Conformal assembly on complex, live biological interfaces. Scanning electron microscopy (SEM) images show the sequential growth of cytoprotective MPN coatings (Eug@MPN₁, Eug@MPN₃) on bare microalgae (Eug), highlighting their interface-sensitive nature [20]. (F) Assembly at soft, fluidic interfaces. A schematic and microscopic images depicting the liquid–liquid phase separation of peptide coacervates stabilized by an ultra-thin MPN membrane, representing the formation of membrane-bound protocells without the need for crystallographic lattice matching [21]. (G) Data-driven optimization via active learning. A closed-loop AI workflow navigates the complex reaction space, identifying target formulations while avoiding redundant precursor paths. The scatter plot demonstrates that active learning strategies can achieve target properties with significantly fewer experimental trials compared to exhaustive sampling [36]. (H,I) Multimodal data integration for AI models. In the absence of standardized crystal structures, machine learning models for dynamic networks rely on heterogeneous inputs and functional readouts, such as (H) X-ray photoelectron spectroscopy (XPS) for local structural fingerprints and (I) target property measurements (e.g., conductivity) for performance optimization [32]. Colors and symbols in panels (G,I) are used for visual differentiation of the schematic pathways, experimental outcomes, and formulation levels. Their local meanings are indicated by the in-panel labels and legends and are not intended to represent a common variable across this figure.

Figure 2. Structural dynamics, interfacial assembly, and data-driven optimization of MPNs. (A) pH-dependent dynamic coordination. Visual color changes and corresponding UV-vis spectra illustrate the transition of local metal–ligand coordination states, demonstrating that MPN structures are highly sensitive to processing conditions rather than adopting fixed periodic lattices [17]. (B) Compositional and morphological versatility. Electron microscopy and energy-dispersive X-ray spectroscopy (EDS) elemental mapping of MPN capsules assembled from various metal ions (e.g., Cu, Al, Zr) [8]. (C,D) Spatiotemporal monitoring of MPNs, showcasing the pathway-dependent nature of dynamic network assembly and structural evolution [41]. (E) Conformal assembly on complex, live biological interfaces. Scanning electron microscopy (SEM) images show the sequential growth of cytoprotective MPN coatings (Eug@MPN₁, Eug@MPN₃) on bare microalgae (Eug), highlighting their interface-sensitive nature [20]. (F) Assembly at soft, fluidic interfaces. A schematic and microscopic images depicting the liquid–liquid phase separation of peptide coacervates stabilized by an ultra-thin MPN membrane, representing the formation of membrane-bound protocells without the need for crystallographic lattice matching [21]. (G) Data-driven optimization via active learning. A closed-loop AI workflow navigates the complex reaction space, identifying target formulations while avoiding redundant precursor paths. The scatter plot demonstrates that active learning strategies can achieve target properties with significantly fewer experimental trials compared to exhaustive sampling [36]. (H,I) Multimodal data integration for AI models. In the absence of standardized crystal structures, machine learning models for dynamic networks rely on heterogeneous inputs and functional readouts, such as (H) X-ray photoelectron spectroscopy (XPS) for local structural fingerprints and (I) target property measurements (e.g., conductivity) for performance optimization [32]. Colors and symbols in panels (G,I) are used for visual differentiation of the schematic pathways, experimental outcomes, and formulation levels. Their local meanings are indicated by the in-panel labels and legends and are not intended to represent a common variable across this figure.

5.1. Encoding Process as Part of Material Identity

For crystalline MOFs, the crystallographic structure often serves as the primary material identity. Although synthesis conditions and defects are important, the periodic framework remains the central descriptor in many AI workflows. For MPNs, this approach is insufficient because the same nominal metal–polyphenol composition can generate different networks under different assembly conditions. Such process and interfacial variables may determine coating thickness, permeability, stability, and release behavior [8]. Therefore, process information should be treated as part of the material identity rather than as secondary metadata. For AI-oriented datasets, each MPN entry should be defined by an integrated record that includes the metal ion, polyphenol ligand, formulation conditions, assembly process, interfacial context, and corresponding functional readout. A useful AI model for dynamic MPNs should integrate component identity, assembly pathway, and application-relevant operating conditions within a unified material representation. This process-encoded representation may also benefit MOF research by improving predictions of synthesizability, defect formation, activation stability, and environmental robustness [42].

Pathway dependence can be handled by representing a dynamic metal–organic material as a time-ordered sequence of evolving material states rather than as a single static structure. At each assembly step, the material state can be encoded by a combination of local coordination graphs, compositional variables, processing conditions, interfacial descriptors, environmental parameters, and intermediate characterization signals. Assembly operations, such as metal-ion or ligand addition, pH adjustment, oxidation, washing, incubation, or substrate exposure, can then be viewed as ordered process actions that transform one material state into the next. Within this framework, graph neural networks can describe local metal–ligand connectivity, whereas temporal or process embeddings can be incorporated through recurrent neural networks, attention-based transformers, temporal graph networks, or neural state-space models to learn how assembly history affects final function. Reinforcement learning and Bayesian optimization can further guide the selection of sequential experimental actions by optimizing reward functions defined by coating stability, permeability, catalytic activity, biological compatibility, or other application-specific outputs. Such architectures would allow AI models to learn not only which composition is optimal, but also which assembly pathway leads to the desired functional material state.

5.2. Learning from Interfaces Rather than Only from Lattices

MOF design has benefited from the ability to extract structural information from periodic lattices. By contrast, MPNs often function at interfaces. They form coatings on particles, proteins, cells, membranes, hydrogels, and porous substrates, and their performance is strongly affected by interfacial chemistry [43]. In such systems, the interface is not merely a boundary condition; it is a central component of material function. Therefore, the interface should be encoded as part of the material state rather than treated as an external substrate. AI-guided MPN design should therefore include descriptors related to substrate type, surface charge, curvature, roughness, hydration, functional groups, and biological or chemical microenvironment. These descriptors can help connect assembly behavior with functional outputs such as adhesion, protection, permeability, catalytic activity, biocompatibility, and controlled release. This interface-sensitive design logic expands the scope of AI from predicting framework properties to optimizing material behavior in realistic environments [44].

5.3. Multimodal Data Integration and Small-Data Strategies

Dynamic metal–organic networks often lack a single structural file that can represent the material. Their structural information is distributed across multiple experimental signals—ranging from spectroscopy and microscopy to scattering and time-resolved measurements. For this reason, multimodal data integration is essential. Rather than relying on crystallographic descriptors alone, AI models should combine spectra, images, process records, and performance readouts to capture local coordination, interfacial coverage, and assembly pathways [23]. Compared with MOFs, MPNs currently lack large standardized databases for such complex inputs. However, this does not mean that AI-guided MPN design must wait until massive datasets are available. A more practical route is to construct small yet information-rich datasets, where each experimental record integrates formulation variables, processing conditions, interfacial context, characterization signals, and functional performance. Active learning, Bayesian optimization, transfer learning, and chemistry-informed models can guide experiments efficiently when the design space is large, but the available data are limited. In this setting, AI does not function only as a predictor trained on existing data; it becomes a partner in navigating a complex experimental design space.

A further bottleneck is that functional datasets for dynamic metal–organic systems are often difficult to compare across interfaces, applications, and readouts. For MPN-type materials, for example, coating stability, permeability, adhesion, catalytic activity, cytoprotection, antioxidant performance, and release behavior may be measured on different substrates, in different media, over different time scales, and using different assay protocols. Similar challenges are expected for related dynamic metal–organic systems whose performance depends on local environment and interfacial state. As a result, even when formulation and process variables are reported, the resulting functional labels may not be directly interchangeable across studies. Future AI-oriented datasets should therefore record not only the numerical performance value, but also the substrate, test environment, assay protocol, normalization method, time point, and uncertainty associated with each functional readout.

5.4. Closed-Loop Experimentation for Dynamic Network Optimization

Closed-loop experimentation provides a practical route to connect crystalline framework design with dynamic network optimization. In MOF research, self-driving laboratories can accelerate synthesis exploration, phase discovery, and performance screening. In MPN research, similar platforms could automate the preparation of metal–polyphenol combinations, control assembly conditions, characterize coating formation, and evaluate function under relevant environments [45]. Such closed-loop workflows are especially valuable for dynamic systems because their performance depends on strongly coupled variables. By integrating automated synthesis, multimodal characterization, machine learning, and functional testing, AI-guided design can move from virtual screening toward autonomous experimental decision-making. The corresponding closed-loop active-learning workflow is summarized in Figure 3, which links automated synthesis, multimodal characterization, small-data modeling, and iterative selection of informative experiments. For MPNs, the closed-loop target should be defined as an application-specific functional state, such as a coating with defined thickness, permeability, stability, and biological or catalytic performance, rather than as a single optimized composition. This transition is central to expanding metal–organic materials design beyond the perfect crystalline lattice.

6. Outlook

The future of AI-guided metal–organic materials design should extend beyond property prediction based on ideal crystalline frameworks and move toward a practical roadmap for dynamic, process-dependent systems. Such a roadmap requires standardized reporting of dynamic metal–organic networks through structured metadata that links composition, assembly conditions, process history, interfacial context, characterization data, and functional readouts. Multimodal characterization protocols are also needed to connect local coordination, network morphology, interfacial coverage, hydration, and pathway-dependent evolution. Function-oriented databases should be organized around application-specific outputs, such as coating stability, permeability, adhesion, catalytic activity, antioxidant capacity, cytoprotection, and release kinetics. Together with active-learning and closed-loop platforms, these resources would enable efficient selection of informative experiments from high-dimensional formulation spaces. In this context, MOFs offer a mature foundation for database construction and structural representation, whereas MPNs reveal the descriptor and experimental challenges posed by dynamic, non-periodic, and interface-driven materials. The broader objective is to generalize topology-centered AI into a materials-state-centered framework that can describe crystalline, defective, amorphous, and adaptive metal–organic systems within a unified design logic.

Building on the logic of curated MOF resources such as CoRE MOF, a minimal metadata schema for MPN-type and related dynamic metal–organic materials should include at least six categories: (i) material composition, including metal identity, valence state, ligand structure, ligand source or purity, and metal-to-ligand ratio; (ii) assembly conditions, including pH, solvent, buffer, ionic strength, temperature, concentration, reaction time, and atmosphere; (iii) process history, including order of addition, incubation, oxidation or reduction steps, aging, washing, purification, drying, activation, and storage; (iv) interfacial context, including substrate type, surface chemistry, charge, curvature, roughness, hydration, and biological or chemical microenvironment; (v) characterization data, including spectroscopy, microscopy, scattering, surface charge, thickness, morphology, coordination fingerprints, and time-resolved measurements where available; and (vi) functional readouts, including stability, permeability, adhesion, catalytic activity, antioxidant activity, cytoprotection, antimicrobial performance, release kinetics, or other application-specific outputs. Compared with crystalline MOFs, for which a crystallographic information file can often serve as a central structural record, FAIR data for dynamic networks face additional challenges because the material identity may depend on the full process history and operating environment. Capturing failed experiments, intermediate states, time-dependent evolution, and multimodal characterization signals will therefore be essential for making MPN-type datasets findable, accessible, interoperable, and reusable.

7. Conclusions

Metal–organic materials occupy a broad structural landscape that extends from long-range ordered crystalline frameworks to dynamic coordination networks. AI has already reshaped the design and screening of MOFs, where machine-readable crystal structures, scalable databases, and relatively stable structure–property relationships support topology-centered prediction. Extending this capability to the broader family of metal–organic materials requires descriptors and models that can operate beyond long-range crystallographic order. MPNs offer one timely and representative platform for this transition, while related amorphous MOFs, coordination polymers, supramolecular metal–ligand networks, and biohybrid metal–organic assemblies present analogous descriptor and workflow challenges because their performance is also strongly affected by local coordination, processing history, interfacial context, and pathway-dependent structural evolution. Using structural order as an organizing principle, this Perspective argues that AI-guided design should advance from predicting idealized crystalline topologies toward optimizing process-dependent material states. Unlike conventional inverse design, which typically searches for compositions or topologies that match target properties from idealized crystal structures, dynamic network optimization treats assembly pathway, interfacial context, operating environment, and functional material state as coupled design variables. This evolution builds on the conceptual and methodological foundation established by MOF-based AI while expanding its reach to less periodic yet highly functional coordination systems. Future AI-guided research should therefore integrate composition, local coordination, processing history, interfacial context, and functional performance into a unified framework for dynamic network optimization.