Dynamical Material Networks Overview

• Dynamical material networks are evolving, graph-based models that encode material structures and their temporal changes across diverse systems such as alloys and granular media.
• They utilize higher-order, multilayer, and compartmental frameworks to represent both static connectivity and adaptive, learning-driven interactions.
• These networks enable predictive insights into alloy discovery, adaptive structural responses, and circular material flows through topology-driven analysis.

Dynamical material networks are network-based representations in which material entities, interactions, or compartments evolve in time. In current research, the expression is used in several technically distinct but related ways: as time-resolved graphs of experimentally known glass-forming alloy systems; as load-bearing structures composed of interconnected adaptable material nodes with trainable physical parameters; as evolving force, contact, or transport networks in granular, elastic, and biological media; and as thermodynamical material networks for circular material flows (Zhang et al., 22 Jul 2025, Kergariou et al., 5 Jun 2026, Papadopoulos et al., 2016, Zocco et al., 2021). This suggests a common core: graph, higher-order, multilayer, or compartmental formalisms are used to encode material structure together with explicit temporal evolution, so that topology becomes a state variable, a design variable, or both.

1. Terminology and conceptual scope

In the amorphous-alloy literature, a material network is a graph-theoretic representation of experimentally known glass-forming alloy systems, and a dynamical material network is a time-resolved version built by accumulating alloy systems up to a given year (Zhang et al., 22 Jul 2025). In the circular-economy literature, a Thermodynamical Material Network (TMN) is “a set $\mathcal{N}$ of connected thermodynamic compartments that transport, store, and transform a target material and whose modeling is based on compartmental dynamical thermodynamics,” with graph vertices representing storage or processing compartments and arcs representing transport compartments (Zocco, 2022). In the structural-materials literature, an Engineering Material Neural Network (EMNN) is defined as a load-bearing architected material structure composed of an assembly of interconnected material nodes with trainable physical parameters and neural-network-inspired morphology (Kergariou et al., 5 Jun 2026).

These definitions are not identical, but they share a formal motif. Nodes may be chemical elements, compartments, particles, beams, or voxels; links may represent alloy co-occurrence, mass transport, physical couplings, or force transmission; and time may enter through discovery year, structural adaptation, loading history, or compartmental dynamics. A plausible implication is that “dynamical material networks” is best understood as a family of representations rather than a single canonical formalism.

The mathematical literature supplies an abstract counterpart. A network can be viewed as a graph $G$ together with a phase-space assignment $\mathcal{P}:G_0\to\mathrm{Man}$, with dynamics constructed from open systems on input trees and interconnected into a vector field on the total phase space $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$ (DeVille et al., 2013). This perspective is sufficiently general to encompass material compartments, interacting microstructural elements, and adaptive physical nodes.

2. Time-resolved material-discovery networks in amorphous alloys

The most explicit use of the term appears in the study of metallic-glass discovery networks. There, the binary material network contains 38 nodes and 94 edges, each edge joining two elements that form at least one binary metallic glass, while the ternary material network contains 47 nodes and 352 ternary glass-forming systems represented as triangles (3-cliques); from these 352 ternary systems, the underlying graph has 348 distinct edges (Zhang et al., 22 Jul 2025). The ternary case is treated as a higher-order network in which the fundamental entities are not only nodes and edges but also triangles, and yearly snapshots are built by attaching an earliest discovery year to each edge or triangle and accumulating systems up to year $t$.

This construction yields a historical map of alloy design. The node count $\mathcal{N}_{\text{node}(t)}$ and the entity counts $\mathcal{A}_2(t)=|E_B(t)|$ for binaries and $\mathcal{A}_3(t)=|\Delta_T(t)|$ for ternaries grow rapidly from approximately 1970 to approximately 1980. After approximately 1990, $\mathcal{N}_{\text{node}(t)}$ nearly saturates, with essentially no new elements introduced into metallic-glass design, and after 2000 the growth of new systems slows, indicating a shift from discovering new systems to exploring properties of existing ones. The innovation classes $B_0,B_1,B_2$ for binaries and $G$0 for ternaries make this explicit: after approximately 1988, almost no brand-new elements appear, and nearly all new systems are $G$1 or $G$2, so nodes are “old” while edges and triangles are “new.” The paper characterizes this as an innovation trap.

The central design mechanism is a topology-based classification of triangles. In a given ternary snapshot, Real triangles correspond to experimentally discovered ternary metallic glasses; Auto triangles are cliques that exist in the element–element graph although no ternary metallic glass has yet been reported; Fake triangles are triples missing at least one edge; and Unknown triangles involve at least one node not yet present in the network. In the full current ternary network there are 836 Auto triangles, 15,027 Fake triangles, and 352 Real triangles; in the full binary network there are 62 Auto triangles and 8,374 Fake triangles. These Auto and Fake sets function as hidden candidate pools encoded by topology rather than by explicit experimental validation.

The temporal statistics are especially strong. Around approximately 1980, Auto overtakes Real, meaning that more potential ternaries are encoded as cliques of known edges than are actually realized experimentally. Most Real ternaries emerged from the Fake category, a smaller but significant fraction emerged from the Auto pool, and only a small fraction came from Unknown. The sharpest retrospective statement is that since 1989, all developed ternary metallic glasses were already encoded in the existing ternary network as Fake or Auto. Cross-layer analysis further shows that among the 352 Real ternary systems, approximately 78% appear as Fake triangles and approximately 11% as Auto triangles in the binary network, which indicates that most ternary metallic glasses could have been anticipated from binary metallic-glass data alone.

Topologically, the binary degree $G$3 and ternary triangle degree $G$4 follow $G$5 with $G$6 for both $G$7 and $G$8, described as abnormal scale-free behavior. Preferential attachment toward hub elements such as Al, Fe, Ni, and Zr is therefore much stronger than in a random graph picture. The authors explicitly propose using Auto and Fake triangles, and higher-order cliques with $G$9, as a topology-driven recommendation system for intelligent amorphous-alloy design.

3. Trainable physical networks and material intelligence

A second research line uses the network concept in an embodied, trainable sense. An Engineering Material Neural Network is defined as a load-bearing architected material structure composed of an assembly of interconnected material nodes, with trainable physical parameters, neural-network-inspired morphology, and learning behavior such that the structure approximates a physical function that maps inputs such as loads, fields, or environmental variables to outputs such as deformation, displacement, temperature, or conductivity (Kergariou et al., 5 Jun 2026). The paper states three defining rules: interconnected adaptable nodes, adaptive output, and energy-efficient, hardware-centric learning.

The analogy to artificial neural networks is direct: neurons correspond to material nodes, weights $\mathcal{P}:G_0\to\mathrm{Man}$0 correspond to physical couplings such as stiffness, conductivity, or magnetization, activations correspond to physical state variables such as displacement $\mathcal{P}:G_0\to\mathrm{Man}$1, strain $\mathcal{P}:G_0\to\mathrm{Man}$2, temperature $\mathcal{P}:G_0\to\mathrm{Man}$3, or electric potential $\mathcal{P}:G_0\to\mathrm{Man}$4, and layers correspond to stacked material layers or hierarchical scales. In the thermal example, the physical mapping is written as

$\mathcal{P}:G_0\to\mathrm{Man}$5

with loss $\mathcal{P}:G_0\to\mathrm{Man}$6 and the conceptual update

$\mathcal{P}:G_0\to\mathrm{Man}$7

What differentiates this framework from digital neural networks is that $\mathcal{P}:G_0\to\mathrm{Man}$8 denotes material properties that must be changed physically, by deformation, phase change, growth, mineralization, jamming, or related mechanisms.

The proposed substrates include composites, architected and microstructured metamaterials, biological and bio-inspired materials, and engineering living materials. The paper also classifies adaptive materials as NPAM, PNRAM, and PRAM, with PRAMs described as ideal for multiple training cycles, analog weight tuning, and reconfiguration. Design and training routes include finite element modeling, voxel-based FE, digital twins, adjoint models, direct feedback alignment, equilibrium propagation, Hamiltonian Echo Backpropagation, Extreme Learning Machines, and Reservoir Computing. This research direction therefore treats the material itself as a computational network whose topology and constitutive parameters both participate in learning.

At the constitutive level, related work embeds neural correction directly inside time-stepping mechanics. In “Neural Material,” the network receives principal stretches and principal stretch rates, $\mathcal{P}:G_0\to\mathrm{Man}$9, and outputs a stress correction $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$0 added to a nominal co-rotational elastic model, so that the total stress is $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$1 (Wang et al., 2018). The model therefore learns nonlinear elastic forces and damping from examples of dynamic behavior of a deformable object’s surface, using sparse reduced space-time optimization to reconstruct internal trajectories and generate training targets. In this setting, a dynamical material network is not merely a graph of parts; it is a dynamic constitutive law embedded inside the simulator.

4. Evolving interaction networks in particulate, elastic, and transport media

A third usage treats material systems as evolving interaction networks whose topology is inferred from physical dynamics. In dense particulate media, the force network at each snapshot is built as a weighted simplicial complex whose vertices are particles and whose edge weights are interparticle force magnitudes; superlevel sets $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$2 then define a filtration over force thresholds (Kramar et al., 2014). Persistent homology of $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$3 and $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$4 yields persistence diagrams for connected components and loops, and temporal change is quantified by bottleneck distance $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$5 and Wasserstein distances $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$6. The main dynamical finding is a three-regime picture: localized large changes in unjammed states, global reorganization near jamming, and much slower evolution in jammed states. Friction strongly affects the topology of both component and loop dynamics.

The same granular setting has also been formulated as a multilayer network. In a quasi-two-dimensional aggregate of 1004 photoelastic disks under biaxial compression, particles are nodes, normal or tangential contact forces are weighted edges, and compression steps form ordered network layers connected by inter-layer identity edges of strength $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$7 (Papadopoulos et al., 2016). Multilayer modularity maximization with a geographical null model yields communities interpreted as evolving mesoscale force-chain-like modules. Node flexibility,

$\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$8

and community stationarity quantify reconfiguration. Normal-force communities are larger, less flexible, and more stationary; tangential-force communities are smaller, denser, and more sensitive to frictional heterogeneity. This provides a direct network-science description of material reorganization under load.

In disordered elastic networks, dynamical data themselves become the training substrate. A graph-neural-network-based simulator trained on uniaxial compression trajectories of two-dimensional disordered elastic networks can reproduce node dynamics and predict emergent properties such as Poisson’s ratio, including cases beyond the training range of Poisson’s ratios (Salman et al., 27 May 2025). The supporting exposition frames the material as a graph of beads connected by harmonic bonds generated from jammed packings, with auxiliary procedures to generate low-Poisson and auxetic networks. A plausible implication is that time-resolved network dynamics furnish a more data-efficient route to structure-to-property learning than static descriptors alone.

Transport on branched biological networks provides an analogous but PDE-governed example. In neurite networks, reaction–diffusion–advection transport is simulated by isogeometric analysis and learned by a graph neural network that decomposes geometry into pipes and bifurcations (Li et al., 2021). The learned simulator predicts dynamical concentration change with an average error less than 10% and 120–330 times faster compared to IGA simulations, while preserving sensitivity to topology, branch radii, curvature, and bifurcation density. Here the network is simultaneously geometric, dynamic, and material: transport paths are encoded as graph structure, and concentration evolution is the state propagated on that graph.

5. Thermodynamical material networks and circular material flows

In the circular-economy and process-systems literature, dynamical material networks are formalized as thermodynamic compartment networks. A TMN is partitioned into $\mathbb{P}G=\prod_{a\in G_0}\mathcal{P}(a)$9, the set of vertex-compartments that store, transform, or use the target material, and $t$0, the set of arc-compartments that transport it (Zocco et al., 2021). Its weighted mass-flow digraph $t$1 uses vertex weights $t$2 for mass stocks and arc weights $t$3 for mass flow rates, and the associated mass-flow matrix is

$t$4

Local mass dynamics obey

$t$5

while energy dynamics obey

$t$6

The paper further shows that Lagrange’s equations can be derived from this dynamical power balance, so mechanical transport devices and chemical reactors can be treated within the same compartmental formalism.

Circularity is then defined in graph terms. A thermodynamically circular flow corresponds to an ordered sequence of compartments $t$7 that begins and ends in the same compartment while processing the relevant material set. In the graph-based formalism, cycles $t$8 in $t$9 are assigned cycle means, and the circularity indicator is

$\mathcal{N}_{\text{node}(t)}$0

with $\mathcal{N}_{\text{node}(t)}$1 the set of non-cyclic flows (Zocco et al., 2021). The 2022 continuation develops several circularity indicators based on geometric, harmonic, and arithmetic means of cycle flows, together with auxiliary quantities such as total stock, total flow, stock distribution, connectivity, cyclicity, directionality, and accumulation–depletion vectors (Zocco, 2022). Dynamic and stochastic versions are also defined by promoting $\mathcal{N}_{\text{node}(t)}$2 to $\mathcal{N}_{\text{node}(t)}$3 or random variables.

The biomethane example illustrates the intended engineering scope. The network contains a biomass hub, a truck compartment, and an anaerobic digestion plant with a reservoir and digester. The truck is modeled via a Lagrangian mechanical subsystem; the digester is modeled by a four-state nonlinear compartmental biochemical system with Monod-type kinetics; and a finite-time stabilizing control law is applied to the digester to drive it to a desired equilibrium. In this setting, planning corresponds to adding, removing, or modifying compartments and arcs, while control corresponds to regulating the dynamical states of the compartments so that the circular material flow remains feasible and efficient.

6. Mathematical foundations, reduction, and adaptive topology

Several mathematical frameworks generalize the notion of dynamical material networks beyond any particular application. The theory of modular dynamical systems on networks models a network as a directed graph $\mathcal{N}_{\text{node}(t)}$4 with a phase-space assignment $\mathcal{N}_{\text{node}(t)}$5, and constructs global dynamics from node-level open systems defined on input trees (DeVille et al., 2013). Graph fibrations then induce maps between network dynamical systems. Surjective graph fibrations yield invariant synchrony subspaces, whereas injective graph fibrations yield surjective maps from large dynamical systems to smaller ones, interpretable as abstractions or fast/slow decompositions. For material networks, this gives a rigorous language for symmetry reduction, coarse-graining, and synchronized substructure.

Spectral reduction offers a complementary route. For node dynamics of the form

$\mathcal{N}_{\text{node}(t)}$6

low-dimensional observables $\mathcal{N}_{\text{node}(t)}$7 can be constructed from dominant eigenvectors of $\mathcal{N}_{\text{node}(t)}$8, leading to effective reduced equations that preserve collective dynamics and critical points (Laurence et al., 2018). The method predicts multiple activation of modular networks and critical points of random networks with arbitrary degree distributions. A plausible implication is that macroscopic material responses in networked microstructures can often be understood through a small number of spectral modes rather than the full state dimension.

The kinetic theory of particles interacting through a dynamical network of links provides a multiscale example in which network remodeling is explicit (Barré et al., 2016). Starting from overdamped particles linked by transient pairwise potentials, the mean-field limit yields coupled kinetic equations for particle and link densities. In the fast-remodeling regime, link density is slaved to particle density, and the system reduces to a single aggregation–diffusion equation. For a Hookean potential, the analysis gives a precise condition for phase transition and uses central manifold reduction to characterize the bifurcation at instability onset. This framework shows how microscopic link turnover can induce macroscopic clustering or homogeneous phases.

Topology can itself become a learning variable. In self-organizing dynamical networks able to learn autonomously, each adjacency entry $\mathcal{N}_{\text{node}(t)}$9 evolves as a stochastic non-Markovian telegraphic signal whose update depends on delayed changes in link state and delayed changes in network error (Kaluza, 2018). Beneficial structural changes are consolidated, harmful ones are reverted, and neutral cases are explored with mutation probability $\mathcal{A}_2(t)=|E_B(t)|$0. Although the paper treats flow-processing networks and first-order chemical reaction systems, the formal mechanism is directly relevant whenever a material network can add or remove couplings as part of its functional adaptation.

Hierarchical deep material networks supply yet another mathematically structured version of the concept. A Deep Material Network is a perfect binary tree of rank-1 laminate homogenization operators, and the micromechanics-informed parametric DMN maps microstructural parameters $\mathcal{A}_2(t)=|E_B(t)|$1 to network weights and rotations, then to effective stiffness, conductivity, and thermal expansion (Li, 2023). Because the same architecture is reused across elasticity, thermal conduction, and thermoelasticity, the network behaves as a morphologically parameterized surrogate for heterogeneous materials with varying microstructure.

7. Limitations, ambiguities, and research directions

The existing literature makes clear that dynamical material networks are powerful but heterogeneous constructs. In the metallic-glass discovery setting, the network deliberately ignores composition details and processing conditions, so it cannot select an optimal composition within a system and is limited by data incompleteness, positive-only data, and a binary/ternary focus (Zhang et al., 22 Jul 2025). In EMNNs and related trainable materials, the major constraints are reliability, degradation, fatigue, controllability of nonlinear dynamics, and the scarcity of Permanently Reversible Adaptive Materials with high cycling stability and sufficiently continuous property ranges (Kergariou et al., 5 Jun 2026). In TMNs, current formalisms emphasize continuous-time and hybrid compartment dynamics, but large-scale optimization, multi-material coupling, and fully exergy-aware circularity metrics remain open (Zocco, 2022, Zocco et al., 2021).

A broader conceptual limitation is terminological. The same phrase can denote a historical graph of alloy discovery, a trainable lattice with physical weights, a force network under compression, or a thermodynamic compartment network. This suggests that future consolidation may depend less on a single definition than on common mathematical primitives: time-indexed graphs, multilayer or higher-order structure, coupled local dynamics, and task- or property-oriented network inference.

Several concrete directions recur across the literature. One is enrichment of state and edge features: composition, processing conditions, and property weights in alloy networks; property-aware weights and composition-resolved edges in TMNs; and explicit sensing–actuation–learning variables in EMNNs. A second is machine-learning integration, especially graph neural networks for link prediction, property prediction, and surrogate simulation in both discovery networks and physically embodied materials. A third is multiscale reduction, where graph fibrations, spectral reduction, and hierarchical laminate networks offer different paths from microscopic interaction networks to macroscopic constitutive behavior. Taken together, these strands indicate that dynamical material networks are becoming a unifying research program for materials whose structure, function, and history are most naturally described as evolving networks rather than as static continua.