Thermodynamic Networks: Theory & Applications

• Thermodynamic networks are systems defined by state variables such as entropy and free energy that capture complex dynamics.
• They integrate graph-theoretic, chemical, and computational frameworks using tools like partition functions, entropy measures, and spectral operators.
• Their applications span event detection in financial and biological systems, molecular computing, and neural network modeling, illustrating interdisciplinary impact.

Thermodynamic networks are networked systems whose admissible states, dynamics, or inferential objectives are formulated through thermodynamic quantities such as entropy, free energy, affinities, partition functions, temperature, or dissipation. In the cited literature, the term spans several distinct constructions: graph-theoretic descriptions in which a network is mapped to a Boltzmann partition function and represented in a thermodynamic state space (Ye et al., 2015); open chemical reaction networks driven by chemostats and analyzed through entropy production, emergent cycles, Gibbs or semigrand Gibbs potentials, and stationary feasibility regions (Polettini et al., 2014, Rao et al., 2016, Liang et al., 2024, Avanzini et al., 2020); molecular-computing models in which thermodynamic favorability is defined by lexicographic optimization of bond count and polymer count (Doty et al., 2017, Breik et al., 2017); and physics-based computational architectures in which relaxation to a non-equilibrium steady state implements a function (Hylton, 2019, Lipka-Bartosik et al., 15 May 2026). A broader physical interpretation also appears in the view that natural networks are open systems whose growth and restructuring are governed by least-time dispersal of free energy (Hartonen et al., 2011).

1. Conceptual scope

A recurring formulation treats nodes as localized thermodynamic units and edges as transport or interaction channels. In autonomous computing networks, the nodes are finite-size reservoirs storing a conserved quantity $q_\alpha$ with conjugate potential $\mu_\alpha$, and the edges carry currents $I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$; conservation then gives $C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$, and the steady state $\mu^\star$ defines the input–output map $N_\theta(x)$ (Lipka-Bartosik et al., 15 May 2026). In a different but related physical picture, natural networks are described as open systems in which nodes are repositories of energy and links are pathways for energy transduction, with entropy production written as

$$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$

and network evolution interpreted as least-time consumption of free energy (Hartonen et al., 2011).

The literature is therefore not unified by a single formalism. It is unified by the decision to describe a network through thermodynamic state variables, constraints, or response functions rather than only through topology or kinetics. In some papers this is a literal nonequilibrium thermodynamics of matter exchange; in others it is an explicitly formal analogy grounded in spectral operators, combinatorial entropy, or steady-state computation.

2. Graph-theoretic thermodynamic descriptions

One major line of work constructs thermodynamic variables directly from graph operators. In "Thermodynamic characterization of networks using graph polynomials" (Ye et al., 2015), an undirected graph $G(V,W)$ is represented by its normalized Laplacian

$\tilde L = D^{-1/2}LD^{-1/2},$

and the network Hamiltonian is chosen as $\hat H=-\tilde L$. The central approximation relates the canonical partition function to the quasi characteristic polynomial,

$\mu_\alpha$0

under the stated conditions $\mu_\alpha$1 and negligible residual term. Average energy and entropy are then approximated without explicit eigendecomposition, using

$\mu_\alpha$2

$\mu_\alpha$3

where $\mu_\alpha$4 is a degree-weighted edge statistic and $\mu_\alpha$5 a degree-weighted triangle statistic. For time-varying networks, temperature is defined isochorically between consecutive graphs by

$\mu_\alpha$6

This yields a trajectory in $\mu_\alpha$7 space, and the paper reports event detection in financial and biological time series, including Black Monday and developmental transitions in a Drosophila melanogaster gene regulatory network (Ye et al., 2015).

A related inference framework uses a Gibbs density matrix on the combinatorial Laplacian,

$\mu_\alpha$8

and fits network models by minimizing the quantum relative entropy

$\mu_\alpha$9

between the empirical density matrix and the model density matrix (Nicolini et al., 2018). Here $I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$0 is interpreted as an inverse temperature, diffusion-time, or resolution parameter, and the fitting objective becomes

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$1

This moves network model fitting away from hand-chosen descriptors and toward an entropy-based spectral comparison (Nicolini et al., 2018).

A third graph-spectral formulation appears in virus–host biomolecular networks represented by a weighted directed adjacency matrix (Rovenchak et al., 20 Jun 2026). There the thermodynamic quantities are derived from adjacency eigenvalues $I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$2, with graph temperature

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$3

partition function

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$4

heat capacity

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$5

magnetization

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$6

and susceptibility

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$7

In Homo sapiens, Mus musculus, and Gallus gallus virus–host networks, targeted removal of influential nodes produces transition-like behavior, including susceptibility inflection points and one- or two-peak heat-capacity curves (Rovenchak et al., 20 Jun 2026).

3. Open chemical reaction networks

In chemical reaction network theory, thermodynamic networks are usually open, driven systems. A general deterministic formulation writes

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$8

with internal species evolving under reaction currents and chemostatted species exchanging matter with reservoirs (Rao et al., 2016). In this setting, the nonequilibrium Gibbs free energy

$I_{\alpha\beta}(\mu_\alpha,\mu_\beta)$9

obeys

$C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$0

and for open detailed-balanced networks the transformed Gibbs potential $C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$1 is the Lyapunov function. For complex-balanced networks, entropy production splits into nonnegative adiabatic and nonadiabatic parts (Rao et al., 2016).

The topological structure of open CRNs is sharpened in the analysis of emergent cycles and broken conservation laws (Polettini et al., 2014). When some species are chemostatted, the internal stoichiometric matrix can acquire new right-null vectors $C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$2 satisfying

$C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$3

which are emergent cycles. Their affinities are

$C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$4

and the steady-state entropy production decomposes as

$C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$5

The paper’s central counting law,

$C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$6

relates the number of independent affinities $C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$7, the number of broken conservation laws $C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$8, and the number of chemostats $C_\alpha \dot\mu_\alpha=-I_\alpha(\mu)$9 (Polettini et al., 2014).

More recent work introduces the thermodynamic space of a CRN as the stationary region of concentration space accessible under finite nonequilibrium driving (Liang et al., 2024). With elementary flux modes $\mu^\star$0 and cycle affinities $\mu^\star$1, individual reaction affinities and concentration monomials are bounded by pathway thermodynamics rather than detailed kinetic solution. The resulting framework yields exact upper and lower bounds on stationary affinities and concentration combinations in examples such as the Schlögl model and a minimal self-assembly process (Liang et al., 2024).

The same thermodynamic structure can be extended to non-elementary effective kinetics (Avanzini et al., 2020). If fast species are coarse grained and the quasi-steady-state current lies in $\mu^\star$2, it decomposes as

$\mu^\star$3

which induces an effective stoichiometric matrix $\mu^\star$4. The effective entropy production remains

$\mu^\star$5

with effective reaction free energy

$\mu^\star$6

For open systems, the semigrand Gibbs free energy satisfies

$\mu^\star$7

(Avanzini et al., 2020).

Two further extensions emphasize thermodynamic ranking and computation at scale. Pathway ranking under fixed throughput currents defines a pathway cost $\mu^\star$8 from the large-deviation rate function $\mu^\star$9, with exact decomposition into maintenance cost $N_\theta(x)$0 and restriction cost $N_\theta(x)$1 (Gagrani et al., 30 Jun 2025). For large stochastic CRNs governed by the chemical master equation, tensor networks make it possible to estimate entropy production rate, heat flux, chemical work, and nonequilibrium free-energy-like quantities directly from the ensemble dynamics, using a Rényi-$N_\theta(x)$2 entropy-rate proxy when the Shannon term is intractable (Nicholson et al., 22 Dec 2025).

4. Thermodynamic binding networks

A distinct meaning of thermodynamic networks appears in molecular computing. A thermodynamic binding network is a geometry-free abstraction in which monomers are multisets of complementary binding sites, and a configuration is a matching among complementary site occurrences (Doty et al., 2017). The thermodynamic objective is lexicographic: $N_\theta(x)$3 A configuration is saturated if no further complementary pairing can be added, and stable iff it is saturated and attains the maximum number of polymers among saturated configurations. In the original formulation, enthalpy is the number of binding edges $N_\theta(x)$4, entropy is the number of connected components $N_\theta(x)$5, and the favored states are those in $N_\theta(x)$6 (Doty et al., 2017).

This framework was used to design Boolean AND/OR formulas and a self-assembling binary counter whose thermodynamically favored states are the desired outputs (Doty et al., 2017). The follow-up complexity analysis shows that deciding whether a TBN has a saturated configuration with at least $N_\theta(x)$7 polymers is NP-complete, while $N_\theta(x)$8 and $N_\theta(x)$9 are $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$0-complete (Breik et al., 2017). The same paper gives a practical SAT encoding using Boolean variables $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$1, $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$2, and representative variables $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$3, and reports that a Boolean sorting-network TBN with about $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$4 saturated configurations can still be analyzed in under a minute (Breik et al., 2017).

The framework was then extended from equilibrium structure to kinetic barriers (Breik et al., 2018). The central claim is that one can program substrate-independent kinetic barriers using only the thermodynamic driving forces of bond formation and configurational entropy. In this setting, a reaction can be thermodynamically favorable yet require crossing a large barrier unless a catalyst is present, and the predicted barriers are robust across variations in the kinetic abstraction (Breik et al., 2018).

5. Learning systems and autonomous computation

Another branch of the literature uses thermodynamic constraints to define or train computational networks. In "Thermodynamic Consistent Neural Networks for Learning Material Interfacial Mechanics" (Zhang et al., 2020), the network input is $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$5 and the outputs are the normal and tangential $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$6-integrals,

$$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$7

Damage variables are defined by

$$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$8

and the loss augments data fitting with three thermodynamic consistency terms: positive energy dissipation, steepest energy dissipation gradient, and energy-conservative loading path. On a silicon/epoxy interface dataset with 10 loading paths and 236 TSR data points, the paper reports thermodynamic violations of about $$\frac{dS}{dt}=k_B L \ge 0,\qquad L=-\sum_{j,k}\frac{dN_j}{dt}\frac{A_{jk}}{k_B T},$$9, compared with $G(V,W)$0 for models without thermodynamic constraints (Zhang et al., 2020).

A more explicitly physical "Thermodynamic Neural Network" treats node states as electrical potentials $G(V,W)$1, edge weights $G(V,W)$2 as transport capacities, and node updates as Boltzmann sampling under a local energy $G(V,W)$3 (Hylton, 2019). The network couples fast reversible equilibration of node states with slow irreversible adaptation of edge states, and interprets self-organization as the result of charge transport, conservation, fluctuation, and dissipation in an open system (Hylton, 2019).

In "Thermodynamic Networks: Harnessing Non-Equilibrium Steady States for Computation" (Lipka-Bartosik et al., 15 May 2026), the computational output is the steady state of a network of finite-size reservoirs. The decisive physical property is Negative Differential Conductance. Without NDC, the conductance matrix remains cooperative and the network computes only monotonic functions; with NDC, the system becomes a universal approximator under the stated assumptions. The framework is instantiated in quantum dot networks and enzymatic reaction networks, and the paper reports successful sine-function approximation and MNIST digit classification, including $G(V,W)$4 test accuracy for an NDC-based network with $G(V,W)$5 hidden nodes (Lipka-Bartosik et al., 15 May 2026).

A related cyber-physical application models power packet routers as macroscopic information ratchets (Hikihara, 28 Mar 2026). A single router is described by

$G(V,W)$6

and the evaluation function balancing gain, information-processing cost, and entropy penalty is

$G(V,W)$7

The paper reports a discontinuous transition at $G(V,W)$8, interpreted as strategic abandon of regulation, and then shows that diffusive coupling in a network of routers extends the bifurcation point and yields collective resilience against local fluctuations (Hikihara, 28 Mar 2026).

6. Recurring principles and limitations

Across these literatures, several principles recur. First, nonequilibrium driving is usually externalized through chemostats, clamped potentials, bias nodes, or fixed throughput currents rather than treated as a spontaneous property of an isolated network (Rao et al., 2016, Liang et al., 2024, Lipka-Bartosik et al., 15 May 2026). Second, topology matters because null spaces, cycles, conservation laws, or conductance matrices determine which thermodynamic forces and steady states are possible (Polettini et al., 2014, Avanzini et al., 2020). Third, thermodynamic state variables often function as compact coarse variables: $G(V,W)$9 trajectories for evolving graphs, semigrand Gibbs potentials for open CRNs, or NESS node potentials for autonomous analog computation (Ye et al., 2015, Rao et al., 2016, Lipka-Bartosik et al., 15 May 2026).

The literature is equally explicit about limits of interpretation. In graph-spectral and biomolecular network analyses, temperature, magnetization, susceptibility, and heat capacity are formal structural observables rather than direct physical temperatures or equilibrium response coefficients of the underlying financial or biological system (Ye et al., 2015, Rovenchak et al., 20 Jun 2026). In thermodynamic binding networks, “enthalpy” is the number of bonds and “entropy” is the number of connected components, with geometry and kinetics intentionally abstracted away (Doty et al., 2017). In TCNNs, the constraints are soft penalties rather than hard guarantees, and TC3 depends on a no-friction dissipation assumption (Zhang et al., 2020). In open-CRN thermodynamic-space analysis, the theory concerns stationary states of ideal dilute, reversible, chemostatted systems and does not directly treat oscillations, chaos, or explicit spatial coupling (Liang et al., 2024).

A plausible synthesis is that thermodynamic networks are best understood not as a single theory but as a family of network formalisms in which thermodynamic reasoning constrains admissible structure, dynamics, or computation. In some cases that reasoning is literal and exact, as in open chemical reaction networks or coarse graining from elementary reactions (Rao et al., 2016, Avanzini et al., 2020). In others it is an intentionally disciplined analogy that turns network spectra, combinatorial connectivity, or steady-state transport laws into thermodynamic state variables and response functions (Nicolini et al., 2018, Lipka-Bartosik et al., 15 May 2026).