WorksPhysics and Astronomy
A scalable computational framework for PDEs on physical networks
Abstract
Transport, reaction and wave propagation on physical networks depend on dynamics along edges and interactions at junctions. We present a flexible framework for modeling physical networks at scale. A network calculus expresses edge dynamics and vertex contributions as fields over the whole network. Differential operators satisfy integration by parts-identities that connect local dynamics to global conservation laws, and explicit vertex constraints allow a variety of physical junction conditions. We implement this formalism using the open-source Dedalus spectral framework. This accommodates heterogeneous edge lengths, material properties and spectral resolutions. Initial-value, boundary-value and eigenvalue problems share a parallel solution method that eliminates internal unknowns independently on subgraphs and couples their solutions through shared boundary vertex interfaces. We verify numerical convergence and demonstrate strong and weak scaling on networks with millions of edges. We present applications to localized optical resonances in photonic networks, preferential channel formation during mineral replacement in evolving pore networks, and nonlinear dispersive waves for KdV on a periodic network.
Introduction
In many applications, the edges of a network represent physical connections with distance along their length. Notable examples include electrical cables, railway lines and blood vessels. Such physical networks, or metric graphs, support differential equations along their edges, coupled by conditions at the vertices.
In this paper we develop a calculus for formulating and solving differential equations on networks as one would on an ordinary continuous domain, such as a torus or disk. Abstract fields describe material properties and physical variables, while differential operators encode both local derivatives and vertex contributions. The gradient and divergence satisfy integration-by-parts identities that connect edge dynamics to conservation across the network. They also satisfy standard calculus rules such as the Leibniz identity. Our numerical implementation is built on the Dedalus Project, an open-source framework for solving differential equations using spectral methods [1]. Earlier developments extended this framework to geometries such as disks and balls [2, 3, 4]. The project combines mathematical structure, flexible model specification and robust numerical algorithms within a shared implementation. Shared algorithms combine testing and maintenance, allowing improvements to benefit many applications. Documented methods and accessible simulation and analysis scripts support scrutiny, reuse and reproducibility [5]. Recent large-scale datasets illustrate the growing scale of whole-network reconstructions: MICrONS maps approximately synapses in a mouse visual cortex [6], Ji et al. reconstructs approximately 6.3 million blood vessel branches per mouse brain [7], and NASA/CNES SWOT provides river geometry and water-storage observations for 126,674 river reaches worldwide, including 21,123 reaches in the Amazon basin [8]. We aim for our infrastructure to enable new network science in a variety of contexts at scale.
We construct a network basis from Dedalus’ cardinal basis for finite sets. Discrete edge, vertex and endpoint labels enter field representations combined with continuous coordinates. Each edge has its own spectral expansion, and the graph’s connectivity is used to prescribe the conditions at vertices. Each physical variable is represented as a single network field, accommodating different edge lengths and material properties. The separation between edge dynamics and vertex coupling enables parallel computation over subgraphs. We partition the network, eliminate internal unknowns independently on each subgraph, and solve a reduced system coupling the interfaces. Partitions balance computational work and reduce communication. The subgraph distribution introduces no additional discretisation approximation.
Related work
Berkolaiko and Kuchment [9] provide the primary reference for the mathematical setting used here, including metric graphs, differential operators, vertex conditions and spectral theory. Earlier surveys describe the basic operator structures and their connections to mathematical physics [10, 11]. A recent computational treatment discusses dynamical processes on metric graphs [12].
Early analytical work establishes spectral characterisations and solvability for network systems [13, 14]. Operator formulations classify self-adjoint and skew-adjoint interactions [15, 16] and extend the analysis to non-self-adjoint problems [17]. Limits of diffusion processes and thin domains provide another route to graph models, deriving effective operators and junction conditions from continuous geometries [18, 19].
Variational and semigroup methods organise diffusion and wave equations within common frameworks [20, 21]. Dynamic vertex conditions couple edge equations to additional evolution at the junctions [22, 23]. These formulations accommodate concentrated storage and other vertex dynamics, as in the heat equation below. General hyperbolic and Hamiltonian systems connect well-posedness to the structure of their energy balances [24, 25]. Higher-order network operators require corresponding extensions of the admissible fields and junction data [26].
Nonlinear dispersive equations form another substantial part of the literature. Studies of nonlinear Schrödinger equations address solitons, standing waves and stability [27, 28], including the dependence of ground-state existence on graph topology [29]. Airy and Korteweg-de Vries (KdV) equations make the relationship between edge orientation, vertex coupling and well-posedness particularly explicit [30, 31, 32]. Related models include travelling waves for the Benjamin–Bona–Mahony equation on networks [33].
Many analytical investigations include computational examples that resolve solution branches and examine stability. Nonlinear Schrödinger equations show how loops and junctions affect bifurcations and localisation [27, 34, 35]. Physical applications include vascular flow [36], gas pipelines and compressors [37, 38], road traffic [39], reactive transport in catalytic networks [40], and network lasers [41]. These examples combine differential equations along edges with junction relations determined by the physical model.
Compatible discretisations preserve relations between fields and differential operators. Finite element exterior calculus connects these relations to topology and numerical stability [42], while mimetic finite differences preserve integral and adjoint identities [43]. Recent examples include twisted and Hessian complexes on polygonal meshes [44], distributional Hessian and div-div complexes containing polynomial fields and Dirac measures [45], and finite-element calculations of spectra and pseudospectra for advection–diffusion of differential forms [46]. § 3.2 gives a simple polynomial example and relates these ideas to spectral discretisation.
Numerical methods address both approximation along individual edges and coupling across the network. Finite-element formulations provide discretisation and error analysis for graph differential operators [47]. Comparisons of spectral, finite-difference and discontinuous Galerkin methods demonstrate different treatments of vertex conditions in eigenvalue, Poisson and wave problems [48]. Further developments include fast graph Fourier transforms [49], domain decomposition for elliptic equations [50], and model reduction that preserves conservation and dissipation properties [51]. Other methods treat fractional elliptic operators [52] and compute nonlinear stationary states through gradient flows [53].
Several software packages provide tools for metric-graph problems. GraFiDi [53] supplies finite-difference operators in Python, with methods for computing ground states and time evolution of nonlinear Schrödinger equations. QGLAB [54] combines finite-difference and Chebyshev discretisations in MATLAB, supporting Poisson problems, eigenvalue calculations, linear and nonlinear time evolution, and continuation of steady solutions through bifurcations. The package netSALT [41] models photonic networks using physically realistic parameters. PETSc DMNetwork [55] provides an environment for parallel multiphysics simulation on networks, including topology and data management, partitioning and distribution, and access to PETSc’s linear, nonlinear and time-integration solvers.
These packages provide precedents for separating graph construction and numerical operators from individual calculations. The present work develops a common calculus of graph fields, differential operators and junction balances, building upon the general Dedalus spectral framework.
An illustrative example
Heat diffusion on a three-edged star graph provides a simple example. The temperature on each edge is
The edges meet at , where the temperature is continuous,
where is the common value at the central vertex. Each edge has a heat capacity per unit length, , and conductivity, . The edge equation is
where the source, , can have distinct values at the junction. Energy balance supplies the vertex equation,
The point capacity, , and source, , allow the junction itself to store and supply heat. Setting both to zero gives the usual Kirchhoff flux condition. The semi-infinite edges isolate the central junction; we assume as . Finite edges instead require conditions at their other ends, determined by the application.
Dedalus can already solve systems assembled in this form, including nonlinear Schrödinger equations on a D-shaped graph [1]. However, specifying each edge equation and junction condition separately becomes cumbersome as the network and model grow. The same heat problem can be written as
Here is a continuous scalar field, the conductivity tensor multiplies the gradient on edge by , and the capacity and source include their point contributions:
On a finite graph, denotes the designated boundary vertices. A homogeneous prescribed temperature condition takes the form . The definitions below make (5) reproduce both the edge equation and the vertex balance.
This example illustrates three advantages of abstraction. First, common operators allow definitions, analysis and numerical methods to be reused across different models and networks. Second, their defining identities connect local dynamics to global balances: the same divergence produces both the edge equations and the vertex flux conditions. Third, a covariant formulation makes additional modelling assumptions explicit. A preferred direction must enter through directional data, and an additional operator or constraint must specify its action on the fields. These requirements provide structural accountability: extensions remain possible, but each must supply the mathematical structure needed to define it and account for its consequences.
KdV, covariance and structural accountability
The linear KdV equation illustrates the need for explicit directional data. The edge equation
changes under reversal of the edge coordinate because
A covariant formulation therefore includes a prescribed direction field. With the unit tangent, , pointing towards increasing , define
The components, , are constant along each edge. Equivalently, for an edgewise arclength field satisfying ; the values of need not agree at vertices. Reversing an edge coordinate changes the signs of both and , leaving unchanged. The formally skew-adjoint directional derivative,
reduces to in each edge interior. Its divergence term supplies the vertex contribution required by the skew identity. The mixed system
defines a differentiated linear KdV model. The real fields, and , are continuous at each vertex. Both equations hold as identities of functionals, with in the second equation interpreted as a regular edge density. Their edge equations are
Thus is constant along each edge. Conditions that set this integration constant to zero recover the linear KdV equation with the prescribed direction.
The local Lagrangian density,
produces (1.11) when varied over the continuous fields. The construction makes the directional field, vertex terms and admissible variations explicit. Establishing well-posed evolution still requires an appropriate choice of endpoint and junction data. Structural accountability exposes those choices and their consequences within the formulation.
Fields and differential operators on a metric graph
Graph fields combine regular functions along edges with data supported at vertices. For scalar fields, these form the hierarchy
Continuous fields share a common value at each vertex. Piecewise fields allow independent incident-edge values, while distributional fields additionally contain Dirac measures and their derivatives. These distinctions specify what a field can represent. Additional conditions, such as a prescribed slope or a relation between values and derivatives, enter as equations.
Several constructions provide precedents. Friedman and Tillich [56] define a graph calculus using edge and vertex measures, with divergence containing both edge derivatives and vertex fluxes. Their measure-valued construction does not supply a hierarchy of derivatives of vertex point masses. Kurasov [57] develops distribution theory for discontinuous test functions at an interface on the line. Nortier [58, §2.3] includes distributional derivatives in a construction for star and rose graphs. Distributional Hessian and div-div complexes on triangulations combine polynomial fields with Dirac measures [45]. Variational formulations of general higher-order network operators describe a broad range of junction conditions [26]. Here we organise regular edge fields, endpoint distributions and constraint operators for direct equation specification and spectral discretisation in Dedalus. Test fields define the operators and their vertex terms; § 3 describes how the resulting equations are enforced using polynomial expansions and tau corrections [59, 60].
Scalar and vector fields
A finite metric graph, , comprises vertices connected by intervals. Each edge, , has length and coordinate . A designated set of vertices, , carries any external boundary conditions. The notation, , counts incident edge ends; a self-loop contributes twice.
Independent of regularity, the symbols, and , denote scalar and tangent-vector field types. The gradient and divergence exchange their tensor character:
Their action on the regularity classes is specified below.
The unit tangent, , points towards increasing . A tangent-vector field has the edge representation, . Reversing the coordinate changes the signs of both the basis and its component, leaving the vector unchanged. Incident edges supply distinct tangent directions at a junction.
Combining regularity with tensor character gives the scalar field classes,
with (c, pw, d) subscripts corresponding to continuous, piecewise and distributional respectively. The class, , contains regular edge fields with independent endpoint values. Regular fields are smooth along each closed edge; their numerical representations are finite polynomial expansions. The class, , additionally identifies all incident scalar values with a common value, . The distributional class, , adds the endpoint Dirac terms defined below. There is no default matching of vector values across a branching vertex. Comparing vectors in different incident tangent lines requires additional structure. Conservation of their summed flux is a separate, intrinsically defined condition.
Each regular edge component has a spectral representation,
where is a degree- basis polynomial. Typically, all edges use the same polynomial family, scaled to their local interval lengths. A continuous-field declaration supplies the endpoint matching automatically. Further restrictions on fields are specified separately as constraints.
The outward derivative of an edge interval is
For edges with distinct endpoints, these are the signs of the oriented incidence matrix, . At a self-loop, the two endpoint signs remain separate in endpoint sums.
Graph integrals use arclength along each edge. The pairing is conjugate-linear in its first argument:
The vector pairing uses the corresponding dot product. Distributional fields act on test fields that are smooth along each closed edge, with independent incident-edge traces. For equations tested against continuous scalar fields, terms acting on endpoint values combine at each vertex because all incident test values coincide.
An endpoint Dirac derivative is defined by
The case, , is an endpoint measure. Higher derivatives act on the corresponding one-sided derivatives of the test field. For continuous scalar tests, all at a vertex act as the same measure, .
A distributional scalar field has the representation
The regular part, , comprises the edge expansions defined above. Only finitely many singular coefficients, , are nonzero. Vector distributions retain the incident-edge tangent basis. A sparse dictionary stores the singular coefficients by vertex, edge end and derivative order. The endpoint labels are part of the distributional field and are retained during multiplication and differentiation. Contributions are combined only when the completed expression is paired with the chosen tests.
Differentiation
The edgewise gradient maps regular scalar fields to regular tangent-vector fields:
It acts in the same way on continuous scalars. For ,
For a piecewise scalar, this gradient differentiates the regular edge components; it does not itself add distributions for differences between endpoint values.
The divergence is defined by its action on all piecewise scalar tests, ,
For a regular vector field, integration by parts gives
Integrated against continuous scalar test functions . Thus
For a continuous scalar , the distributional Laplacian has the following representation when evaluated on continuous scalar tests:
Piecewise regular coefficients multiply endpoint distributions through the product rule:
Expressions and equations
There is an important subtlety in evaluating Dirac distributions at a vertex. A piecewise test function can have different endpoint values on its incident edges, whereas a continuous test function has one common value. For example, at a vertex, , joining edges 1 and 2,
If is continuous at , the result becomes
The two endpoint contributions then act as a single vertex measure, and cancel if . Multiplication by a discontinuous coefficient can weight these contributions differently and remove that cancellation. We therefore retain the individual endpoint distributions throughout the algebra and combine their contributions when the complete equation is evaluated on continuous tests.
With the multiplication defined above, the divergence satisfies the Leibniz rule,
for and . This identity holds before combining endpoint contributions. For discontinuous coefficients, multiplication must precede combination of endpoint contributions. For a continuous test field, ,
Dedalus performs the final combination assuming continuous test fields when assembling the complete equations into the system of spectral coefficients and tau corrections. Continuity of test values does not require continuity of material coefficients or identify incident-edge derivative traces. We leave this final assembly convention implicit in our scalar equations, using for individual endpoint contributions and for the action on continuous tests.
Operator composition
We demonstrate how higher-order derivatives work via a case study of the beam equation. Each edge has a one-dimensional tangent space. An intrinsic tensor of order has the form
Contractions remove pairs of tangent factors because . Even-order tensors therefore reduce to scalars and odd-order tensors to vectors, without losing any edge components. More specifically, all intrinsic rank-2 tensors act by scalar multiplication,
For a piecewise regular tangent-vector field , the edgewise gradient is
The scalar representation of the edgewise Hessian is
The operator, , contains only regular curvature; contains additional vertex terms in (27).
A displacement, , has the beam Lagrangian density
Here is the applied force density, while are the linear density and bending stiffness, respectively. Variation of the action gives
Within an edge,
The adjoint is defined on piecewise scalar tests by
is the bending moment. Two integrations by parts give
The edge expression is fixed. Relations between admissible endpoint variations determine which singular contributions remain independent. For a continuous field with no further constraints, the common vertex value of and its independent incident-edge derivatives give, in the absence of vertex forcing,
Other junction conditions can be imposed by adding explicit constraints and their unknown reaction sources.
The endpoint trace operators provide the missing piece,
with adjoints defined using the standard Hermitian pairing on endpoint arrays,
A linear constraint is specified by
The matrices, , select and combine endpoint data. They can impose a rule throughout the graph or on a designated set of vertices. The constraint rows are chosen independently on the declared field class; continuity already supplied by is not imposed again.
The constrained beam system is
The purpose of this example is to show how explicit constraints supplement the conditions supplied by an adjoint. With the field class fixed, the adjoint gives natural endpoint and junction balances, just as divergence gives Kirchhoff’s law in the heat equation. To impose a different condition, e.g. [61], supplies the constraint and supplies an unknown reaction, with “slack variable” , in the field equation. This is very analogous to the tau method approach used in Dedalus to enforce strong-form systems of equations. The endpoint equations then determine the reaction in place of the corresponding natural condition, while the remaining balances still hold. For example, and prescribe zero slopes while allowing nonzero endpoint moments. The same ideas allow more general options in the KdV example as well.
Table 1 gives representative endpoint and junction conditions obtained by choosing in (46) [26, 62]. All use the same edge operator and distributional adjoint.
| Homogeneous endpoint or vertex conditions | Constraint expression, |
| Prescribed endpoint value: , | |
| Prescribed endpoint slope: , | |
| Mixed endpoint condition, with real : , | |
| Coupled junction: ; all equal; |
Table 1. Examples obtained by setting the indicated constraint expression to zero and including its adjoint reaction in (46). The displacement is continuous and has no vertex-supported terms. The first three rows apply at degree-one endpoints; the last applies at a junction.
Computational discretization
The graph calculus forms an interface layer over the existing Dedalus infrastructure [1]. Users declare graph fields and enter equations in the abstract forms illustrated by the heat, KdV and beam examples. The interface translates these expressions and their trace constraints into edge and vertex equations, then assembles the polynomial coefficients, vertex unknowns and tau corrections. These models use the standard Dedalus problem classes: initial-value problems, linear and nonlinear boundary-value problems, and linear eigenvalue problems. Dedalus supplies the time integrators, Newton iterations and eigenvalue algorithms, while the graph interface supplies the fields, operators and vertex coupling. For linear eigenmodes of a time-independent system, the ansatz, , replaces . Reducing higher-order time derivatives to first order where needed gives Dedalus’s standard generalized eigenvalue form.
The code boxes below illustrate the interface for the static beam problem in § 2.3, using supplied network data. Three unit-length beams have unit bending stiffness, unequal distributed loads, and clamped outer ends. The central displacement is continuous and the outward slopes sum to zero, using the coupled-junction condition in Table 1. All quantities are nondimensional. Each box introduces the code for the corresponding concept.
Network bases and domains
The cardinal basis allows discrete labels to participate in the same field representation as continuous coordinates. For the finite edge set , define
Let denote a reference coordinate on each edge, and take a common polynomial interval basis . The product expansion
represents the edge restrictions of a physical field together as a single Dedalus field, indexed by the edge label and the interval coordinate. Each edge retains its own polynomial coefficients . The physical coordinate on edge is , so
These length factors also enter the evaluation of endpoint derivatives.
The representation on the finite set is exact: a function on has its values as cardinal coefficients, so the transform along this direction is the identity. The polynomial truncation acts along the interval coordinate. The identity makes multiplication by an edge-dependent scalar diagonal in the cardinal direction, while differentiation and polynomial transforms act along each interval. Edge-dependent lengths and material properties therefore fit directly into the usual coefficient operations.
Analogous cardinal spaces index vertices and incident endpoints. Sparse maps between these spaces encode incidence, restriction and sums over incident edges. Together with endpoint evaluation of the interval expansions, these maps connect edge fields to vertex data. The declared field regularity and the supplied vertex equations specify continuity and other coupling conditions. The NetworkBasis interface packages the cardinal and polynomial bases for use with ordinary Dedalus fields and problem classes.
This construction applies to any finite index set; networks supply a particular collection of spaces and maps. It brings discrete coupling into Dedalus’s existing symbolic equations, spectral operators and distributed field storage.
The supplied vertex_ids form an ordered table of vertex labels; endpoints lists ordered pairs of indices into that table, and lengths gives the corresponding edge lengths. The mapping vertex_groups assigns names to sets of vertex-table indices, using clamped for the supports and coupled for the junctions in the beam example. Group names alone impose no conditions. The distributor manages field storage, and the network basis combines the edge cardinal basis with retained polynomial modes per edge. In the code examples, d3 denotes the Dedalus public interface.
network = d3.Network(vertex_ids=vertex_ids, endpoints=endpoints, lengths=lengths,
vertex_groups=vertex_groups)
dist = d3.Distributor(network, dtype=np.float64)
basis = d3.NetworkBasis(network, size=N)
Cohomology and spectral discretisation
The polynomial representation on a graph satisfies the underlying cohomology exactly. That is, take an edgewise polynomial in ,
where is the degree- polynomial basis element on . There is no nontrivial intrinsic local curl operator on a graph. However, still has a Hodge decomposition of the form,
where ,
which is unique up to a global gauge constant . Both expansions use the same local polynomial basis without loss of generality. The remaining field, , is constant on each edge and satisfies the Kirchhoff condition at every vertex, i.e.,
The field, , is orthogonal to the gradient of every continuous scalar,
for all , and carries all the circulation of . That is, for any closed loop, ,
A graph with a single connected component contains linearly independent cycles,
where is a well-defined linearly independent basis of cycles. Expressing vectors and scalars in different polynomial bases changes their coefficients but leaves the cohomology unchanged. The point here is that finite element and spectral implementations share the same compatible polynomial spaces. Dedalus enforces the resulting strong-form edge equations with explicit tau corrections in spectral coefficient space, combined with various vertex equations.
Fields
An ordinary Dedalus field on the network basis stores the displacement on all edges. Declaring regularity=‘continuous’ identifies its incident endpoint values at each vertex; it leaves the slopes independent. The bending stiffness, called EI in the script, is constant. The load uses only basis.edges, so it has one value per edge and is constant along each interval. The local cardinal grid supplies the edge indices 0, 1, and 2, giving loads 24, 48, and 72. The ['g'] accessor sets a field’s grid values. We also retain the named vertex groups for the constraints below.
# Displacement and coefficients
u = dist.Field(name='u', bases=basis, regularity='continuous')
u['g'] = 0
clamped = network.vertex_groups['clamped']
selected_junctions = network.vertex_groups['coupled']
EI = 1.
f = dist.Field(name='f', bases=basis.edges)
edge = dist.local_grid(basis.edges)
f['g'] = 24 * (edge + 1)
Equations and vertex constraints
We specify the static beam balance as a linear boundary-value problem. The equation namespace supplies the fields and coefficients, while H and H.T are available directly in equation strings. They denote the edge curvature and its distributional transpose, including the endpoint moments and forces in (41).
# Problem
problem = d3.LBVP([u], namespace=locals())
problem.add_equation('H.T(EI*H(u)) = f')
The support and joint kinematics are supplied through endpoint, the trace operator in (43). Order deriv=0 gives endpoint values and deriv=1 gives outward physical slopes. The argument on selects vertices; sum=True sums incident endpoint contributions separately at each selected vertex. Thus the first two calls below clamp the outer ends, while the third imposes the chosen junction slope sum. Each add_vertex_constraint call both imposes and introduces its unknown reaction into the beam balance, as in (46).
# Vertex constraints add their unknown adjoint reactions to the beam balance.
problem.add_vertex_constraint('endpoint(u, deriv=0, on=clamped) = 0')
problem.add_vertex_constraint('endpoint(u, deriv=1, on=clamped) = 0')
problem.add_vertex_constraint('endpoint(u, deriv=1, on=selected_junctions, sum=True) = 0')
The equivalent call is problem.add_equation(..., vertex_constraint=True). Edge tau corrections are generated automatically and remain accessible through problem.tau_fields; custom explicit formulations can disable this generation with tau=False in add_equation.
The assembled system
Dedalus discretizes the edge equations using its spectral tau method [1]. Unknown tau coefficients multiply prescribed polynomial corrections and are solved for alongside the field coefficients, allowing junctions to be imposed on the truncated expansions.
For a linear solve, we let collect the expansion and tau coefficients of all fields on edge . collects the additional unknowns associated with vertex conditions and global constraints, including shared vertex values and constraint reactions. Writing , the assembled equations have the block form
The first block row contains the edge equations and endpoint continuity conditions. Each acts on the coefficients belonging to one edge, while connects its endpoint values to the shared vertex unknowns. The second block row imposes vertex balances, prescribed constraints and any global conditions. The block contains contributions from the edge unknowns and contains the coefficients from the vertex, reaction and global unknowns in . The right-hand sides contains the known terms for the given linear solve. The block-diagonal structure of follows from coupling distinct edges only through these retained unknowns.
For the beam example, the assembled system includes the fourth-order edge equations and the unknown support and junction reactions. The usual Dedalus solver interface builds and solves this system:
# Solver
solver = problem.build_solver()
solver.solve()
The chosen loads admit a polynomial exact solution, providing a check of the assembled beam problem. Here is the physical coordinate measured from the central junction along each unit edge. The final box reports the maximum displacement error on the grid for this three-edge example.
# Check against the exact solution
x = basis.local_physical_grid(dist)
exact = 2*(1-x*x)**2 + (edge-1)*(x**4 - 1.5*x**3 + 0.5*x)
error = np.max(np.abs(u['g'] - exact))
print(f'Maximum beam displacement error: {error:.6e}')
Parallelization over subgraphs
We distribute the network over subgraphs and reduce the assembled equations to a coupled system on their interfaces. After solving this system, the eliminated unknowns are recovered locally within each subgraph.
Network partitioning
Subgraphs and interface unknowns
We partition the edge set into disjoint subsets,
and let contain the endpoints of the edges in . The subgraph is assigned to one parallel process, which holds the coefficients and equations associated with its edges. Each edge belongs to exactly one subgraph, while vertices may be shared by several subgraphs. A subgraph may contain many edges and does not need to be connected.
A vertex in is internal to subgraph if all its incident edges belong to ; denote the set of these vertices by . Vertices incident to edges in more than one subgraph form the interface , with the interface of subgraph .
If each junction carries several unknowns—for example, the concentrations of two reacting species—the interface retains whichever of those unknowns couple adjacent subgraphs. Those are determined by the vertex conditions, so they are not automatically common endpoint values. A global constraint can also couple otherwise separate subgraphs. For example, prescribing a network-wide mean may introduce a single multiplier that affects equations throughout the network. That multiplier stays in the reduced system, and each subgraph contributes to the global constraint.
Work estimates
We estimate local work by assigning a weight to each edge. Equal weights balance edge counts and are appropriate when the equations and resolutions are comparable across edges. For heterogeneous discretizations, the number of expansion and tau coefficients provides an inexpensive initial estimate of edge work. For the two-field, second-order problems considered below, an edge with expansion coefficients per field has weight , including its tau coefficients. Length weights are useful when the resolution is approximately proportional to edge length, as in the heat-conduction examples below.
We also assign weights to junctions, estimating the number of scalar interface unknowns associated with each junction. For the same two-field problems, each junction has weight two.
These estimates do not account fully for the cost of factoring the coupled subgraph systems, which also depends on their connectivity and equation structure. A smaller interface can increase local work imbalance or factorization cost, so interface size alone is insufficient to assess a partition. We therefore consider the complete solver cost, including setup and repeated solves, when choosing a partition for an application. Whole edges remain indivisible, which can limit the achievable balance when a few edges dominate the work.
Partition construction
For general networks, we use Mt-KaHyPar (Multi-Threaded Karlsruhe Hypergraph Partitioner) [63] to construct a -subgraph partition. We construct a hypergraph in which the network edges become vertices and the network junctions become hyperedges. For example, three edges meeting at a junction become three vertices joined by one hyperedge. Mt-KaHyPar divides these vertices into groups, balancing their total weight while penalizing hyperedges that cross between groups. In the original network, this balances the work across processes while reducing the number of processes sharing each junction. We use the weighted connectivity objective: a junction contained within one subgraph contributes nothing, while a junction shared by two subgraphs contributes its weight, one shared by three contributes twice its weight, and so on.
If a spatial embedding of the network is provided, we also offer spatial partitioning, which divides the network geometrically while balancing edge weights. Other available constructions grow subgraphs through adjacent edges or assign edges solely to balance work. Users can instead supply their own edge assignments. Once selected, the assignment remains fixed throughout the solve.
Subgraph interior elimination
We collect the edge expansion, tau coefficients and the internal vertex and reaction unknowns of subgraph into . Let contain the retained interface unknowns and any global unknowns. Reordering the assembled equations gives
where . Each block contains the interior equations and unknowns of one subgraph. The remaining blocks couple these interiors to the retained unknowns and impose the interface conditions.
Assuming each is invertible, eliminating the interior unknowns yields
We factor each subgraph matrix as a single sparse system containing the edge coefficients, tau variables and internal vertex and reaction unknowns. These factorizations are computed independently across processes and reused to construct the interface system and reconstruct the local solution. Only the retained unknowns enter the coupled interface solve. The inverse notation denotes applications of local factorizations.
Interface solution and local reconstruction
We assemble the interface system by combining the subgraph contributions to the interface conditions and any global constraints. Each interface equation is assigned to one process, which collects the contributions to that matrix row.
We solve this system using restarted GMRES with diagonal right preconditioning. Each process stores its assigned matrix rows and the corresponding entries of the Krylov vectors. To evaluate a matrix–vector product, it obtains the additional vector entries required by those rows from the processes holding them. Inner products and residual norms are computed by summing contributions across processes. We form the preconditioner from the assembled interface diagonal. We accept an iterative solution when the explicitly evaluated residual satisfies
where and are the absolute and relative solver tolerances. The interior unknowns are then recovered independently through
where and are the block rows associated with subgraph . Only retained unknowns coupled to that subgraph contribute to . Reconstruction recovers both the edge fields and the internal vertex and reaction unknowns, using the local factorizations already computed during elimination.
This reduction and reconstruction are algebraically equivalent to solving the full assembled system whenever the required inverses exist. They introduce no additional discretization approximation; numerical differences arise from roundoff and the tolerances used in the linear solves.
Application to different problem classes
We use Dedalus’s time integration, Newton and eigenvalue algorithms [1], applying the subgraph reduction to their linear solves. The reduction acts on the assembled system for steady linear problems, the Jacobian for Newton corrections, and the shifted systems used in shift-invert eigenvalue calculations. Newton corrections include the linearized vertex conditions.
For initial value problems, we reduce the implicit system at each timestep or stage. Edge and vertex time-derivative terms enter the system together, while algebraic vertex conditions remain constraints.
We reuse local factorizations and the reduced interface operator while the system matrix remains unchanged. When several Runge–Kutta stages have the same implicit matrix, they also share these factorizations and the interface operator. This avoids storing and constructing separate copies for each stage. Changes to the timestep, implicit coefficients, Newton iterate or eigenvalue shift may require rebuilding them.
Numerical verification and parallel performance
We now assess the numerical accuracy and parallel performance of the network solver. We measure convergence against analytical and refined numerical reference solutions, and compare distributed solutions with direct solutions of the assembled system. Residuals of the edge equations and vertex conditions provide an additional check that the computed fields satisfy the prescribed problem. We then measure strong and weak scaling on networks with contrasting connectivity.
Heat conduction in a network of rods
We first consider steady heat conduction through a network of rods that exchange heat with their surroundings. The field denotes temperature relative to a uniform ambient temperature. Each rod has constant coefficients and , describing axial heat conduction and linear heat loss, respectively, and satisfies
We impose continuity of temperature at every vertex and conservation of heat flux at vertices where the temperature is not prescribed:
For an edge with endpoints and (corresponding to and , respectively), we write and the exact temperature profile is
Substituting its endpoint derivatives into the junction balances gives a linear system for the unknown vertex temperatures. Together with the prescribed temperatures, this determines a reference solution independently of the spectral discretization.
We arrange the rods in a square grid with four vertices in each direction, giving 16 vertices and 24 rods. In dimensionless variables, each rod has length and coefficients . We hold the bottom-left vertex at temperature relative to the ambient temperature and impose heat-flux balance at every other vertex. Heat is conducted along the rods and lost continuously to the surroundings. The reference solution follows from a linear system for the 15 unknown vertex temperatures. We vary the number of Chebyshev modes on each rod from to and measure the relative error over the whole network, normalized by the reference solution’s norm. Figure 1(a,b) shows the temperature decaying away from the heated corner and the spatial convergence of the solution. The error falls from approximately at to at , reaching roundoff levels by . We repeat the calculation on two and four processes, using a spatial partitioner with equal weights for all rods. The distributed solutions agree with a direct solve of the assembled system to roundoff at every resolution tested.
Figure 1. Steady and transient heat conduction in networks with 16 vertices, 24 rods and equal total length. (a,b) Steady temperature and relative error against the analytical solution for the uniform grid with a heated bottom-left corner. (c) Transient temperature at with . (d) Relative error at against the reference, with throughout the spatial refinement. Both temperature fields use , giving 768 edge coefficients per network; is the approximate number of modes per unit length. Both colour scales are logarithmic, with different lower limits. Convergence curves compare a direct solve with the subgraph method on two and four processes; the miniature networks show the corresponding partitions, coloured by owning process.
We next consider transient heating on a heterogeneous network with the same 16 vertices, 24 rods and total rod length of 24. We perturb the grid vertices, retain 24 edges from their triangulation while preserving connectivity, and rescale the resulting rod lengths to match the grid’s total length. Each rod satisfies
with constant coefficients on each rod, varying across the network within , and . Starting at ambient temperature, we prescribe at the bottom-left vertex and impose the same continuity and heat-flux conditions at all other vertices. We subtract the instantaneous steady temperature profile before time stepping, giving homogeneous vertex conditions and a known forcing term in the edge equations, and restore it when reconstructing the temperature. We use Dedalus’s second-order RK222 scheme and evaluate the solution at .
Here denotes the nominal number of modes per unit length. We assign integer mode counts approximately in proportion to rod length, with at least four modes per rod and exactly edge coefficients in total, matching the uniform grid at each resolution. The spatial partitioner uses rod lengths as weights. Figure 1(c) shows the transient temperature for with . Figure 1(d) shows spatial convergence for to at fixed . We compare each solution with a direct calculation at the same timestep, so these errors measure spatial refinement at fixed , rather than the total error relative to the continuous-time solution. The curves reach a roundoff-level plateau near ; the and references differ by approximately in relative norm. The direct, two-process and four-process curves agree throughout the refinement range.
A separate timestep-refinement study at fixed nominal resolution , with halved from to , reduces the relative error at by approximately a factor of four per refinement, consistent with second-order time integration. The reference uses and .
Benchmark problems and measurement procedure
We test parallel performance on square grids, branching networks and triangular meshes. The branching networks contain groups of interconnected vertices, each with a highly connected hub, joined by relatively few edges. The triangular meshes are periodic square grids with one randomly oriented diagonal added to each cell.
On every network, we evolve two coupled reaction–diffusion fields:
The coefficients are constant along each edge and vary between edges and fields, with , and . Both fields are continuous at vertices and satisfy conductivity-weighted flux balance, without vertex storage or reaction. The forcing is chosen to provide a known time-dependent solution for checking numerical accuracy.
Edge lengths are sampled uniformly from . We assign approximately 32 Chebyshev modes per unit length to each field, rounded to multiples of eight, giving 16–64 modes per field on each edge. Reported coefficient counts include both edge fields and exclude tau variables and vertex unknowns. We partition the networks using Mt-KaHyPar with coefficient counts as edge weights.
All measurements use AMD Genoa nodes and Dedalus’s RK222 scheme with . Each configuration is run once, averaging the wall time over 1,000 timesteps after the initial steps. Timings include equation evaluation, linear solves and communication, but exclude network construction, partitioning, assembly, initial factorization, validation and output.
Scaling
For strong scaling, we hold the network, coefficients and discretization fixed while increasing the number of processes. We use approximately four million coefficients (50,000 edges) within one compute node and 512 million coefficients (6.4 million edges) across compute nodes. Figure 2(a,Figure 2b) shows speedup relative to one process and one full node, respectively. Within-compute node tests use 1–64 processes; across-node tests use 1–16 nodes with 96 processes per node.
Figure 2. Strong and weak scaling on three network families for equations (68) and (69). (a) Within-compute node speedup for approximately four million coefficients, relative to one process. (b) Across-node speedup for approximately 512 million coefficients, relative to one node with 96 processes. Dotted grey lines indicate ideal speedup. (c) Time per timestep at four approximate coefficient counts per process, as indicated in the legend; ideal weak scaling is horizontal. The vertical line marks one full node. Each point averages 1,000 timesteps.
The three networks achieve speedups of 19–21 on 64 processes. Across compute nodes, scaling is approximately ideal through four nodes, reaching speedups of 12.6–13.2 on sixteen nodes.
For weak scaling, we increase the number of edges while holding the target coefficient count per process fixed. The edge-length distribution, resolution rule and equations remain unchanged. The four workloads span approximately 300–5,000 edges per process; small differences from the target sizes arise from integer network dimensions. Figure 2(c) shows the measured time per timestep, which would remain constant under ideal weak scaling.
At the largest workload, timestep time increases by a factor of 2.4–2.9 between one and 64 processes. Subsequent scaling across fully occupied nodes is substantially better: increasing from one to sixteen compute nodes increases timestep time by 29–35%, while the total problem size grows approximately sixteenfold.
Examples
EVP: passive optical resonances in photonic networks
Model and coupling conditions
Following Saxena et al. [41], we consider passive optical resonances in networks of waveguides constructed from random lines within a square. Each edge has a constant, possibly complex refractive index , accounting for absorption. The degree-one boundary vertices continue into semi-infinite outgoing leads with the same refractive index.
With time dependence , we seek complex resonant wavenumbers , where is the speed of light, and a continuous electric-field amplitude . For , introduce a regular tangent-vector field , with . The network equations are
where
represents the outgoing leads at the open ends , and is the terminal edge incident to $v. The balance equation is evaluated against continuous scalar test fields.
In each edge interior, these equations give the Helmholtz equation
At internal vertices, the divergence supplies Kirchhoff coupling,
where is the outward derivative defined in (18). At an open end, the boundary measure instead gives
imposing an outgoing wave in the lead. The complex resonances account for absorption along the waveguides and radiation through the open ends: gives the oscillation frequency and the amplitude decay rate. The first-order network equations depend linearly on , yielding a generalized eigenvalue problem after spatial discretization.
Reproduction of the reference result
We use the network supplied with Saxena et al. [41], comprising 96 vertices and 131 edges, with a total internal waveguide length of 2500 . The refractive index is on terminal edges and their leads, and on all remaining edges. We seek passive resonances in the spectral window
with measured in . The reference paper reports 454 passive modes in this window. For the numerical solution, we choose a fixed carrier wavenumber near each spectral target and write
This exact change of variables gives
We expand the amplitudes in Chebyshev polynomials and impose the vertex conditions on the reconstructed physical fields, obtaining a linear EVP for . We use 24 Chebyshev coefficients per amplitude on each edge and repeat the calculation with 32 coefficients to check spatial convergence.
We compare the computed spectrum with a fresh run of the archived netSALT implementation supplied with Saxena et al. [41]. For each matched resonance, we compare the complex wavenumber and spatial field profile, accounting for the arbitrary complex amplitude of an eigenmode. An independent calculation based on exact travelling-wave solutions on each edge provides a further check. We recover all 454 resonances in the prescribed spectral window, with each mode matched to the archived netSALT rerun and the independent travelling-wave calculation. Against the latter, the maximum relative differences in and are below and , respectively. The matched spatial profiles agree to floating-point roundoff in normalized overlap.
Larger networks
To assess computational reach, we generate independent networks of the same construction with approximately , , and physical edges. We enlarge the physical domain while keeping the local line density, short-distance cutoff, refractive indices and vertex conditions fixed. At each size, we target 16 resonances near each of six prescribed complex shifts within the reference spectral window, removing duplicates between searches.
At the largest size, we compute 96 distinct resonances on each of three networks containing between 100,410 and 100,676 physical edges. Each mode is checked under spatial refinement and enlargement of the eigensolver subspace. These are selected resonances near the prescribed spectral targets, rather than complete spectra. We quantify localization using the integrated intensity
Figure 3 shows an interior localized mode on the 100,485-edge network. Half of this intensity lies within a 0.599 mm square in the 6.16 mm-wide domain, while the resolved edge profile exhibits an intensity oscillation period of 0.196 m.
Figure 3. Interior passive resonance of a 100,485-edge network. (a) Pointwise intensity , normalized by its maximum over internal edges, . Nested squares share the internal intensity centroid and contain 50%, 90% and 99% of the integrated internal intensity, including partial-edge contributions at their boundaries. (b) The 50% square, of width 598.6 m. Both maps use the same logarithmic colour scale, with values below mapped to its lower endpoint; unresolved optical oscillations are represented by spatially averaged pointwise colours. (c) Resolved intensity over the first 3 m of the circled edge , which contains the internal intensity maximum. The trace uses the same normalization, with a linear vertical axis and no display floor; its oscillation period is 0.1961 m. The resonance has and is unchanged to numerical precision under refinement from 24 to 32 coefficients per envelope.
LBVPs on an evolving network: mineral replacement in porous media
Model and coupling conditions
We use the mineral-replacement model of Budek and Szymczak [64] to demonstrate successive linear boundary-value solves coupled to evolving pore geometry. Each pore has length and uniform diameter . Flow and solute transport are quasisteady relative to mineral evolution.
For a continuous pressure field , the volumetric flow , with , satisfies
where is the fluid viscosity. The divergence balances are evaluated against continuous test fields that vanish at the inlet and outlet reservoirs. Their edge and vertex contributions give constant flow along each pore and conservation of flow at internal junctions. The outlet pressure is zero, and a common inlet pressure is adjusted to maintain total flow . Transverse boundaries are periodic.
The reagent and intermediate concentrations are piecewise scalar fields, using the stoichiometric rescaling of Budek and Szymczak [64]. Their steady transport satisfies
The effective reaction coefficients include diffusion to the pore walls; axial diffusion is neglected. Dissolution stops when a pore has no remaining access to primary mineral.
The vertex terms conserve each species’ advective flux. We complete the coupling by imposing complete mixing: all outgoing edges receive a common concentration, while incoming concentrations may differ. Writing , this gives
where and denote incoming and outgoing edges. At inlet vertices, we prescribe . Outflow boundaries require no concentration condition.
The uniform pore diameter evolves according to
where is the stoichiometrically weighted molar volume of the primary mineral and is the corresponding ratio of precipitated to dissolved solid volumes. The integrated reaction rates also update the primary and secondary mineral inventories in adjacent grains. Pore lengths follow the cube-root grain-volume relation of Budek and Szymczak [64, Eq. (20)], and pores below the prescribed closure diameter are removed from the active network.
At each geometry state, Dedalus solves the pressure LBVP and then the coupled transport LBVP with the flow held fixed. The resulting reaction rates advance the mineral inventories and pore geometry before the next pair of solves. The complete evolution is nonlinear, while both spatial solves are linear.
Reproduction of the reference result
We consider the channeling regime with , and on an independently generated random network of 29 996 pores. Figure 4(a–d) shows channels of enlarged pores extending from the inlet towards the outlet. Dissolution increases the hydraulic conductance of these paths. More flow and reagent is directed along these paths, reinforcing the dissolution process. The channels grow unevenly, with a few extending through most of the network by the final displayed state. The permeability increases throughout the simulation, with a sharper rise as the channels approach the outlet (Figure 4e). This evolution qualitatively reproduces the preferential channel formation and permeability increase reported by Budek and Szymczak [64].
Figure 4. Mineral replacement in the channeling regime (, , ) on an independently generated random network of 29 996 pores. (a–d) Pore diameters at the indicated times, with flow from top to bottom and periodic transverse boundaries. Colours show on a shared linear scale, where is the uniform initial diameter; thicker lines indicate wider pores. (e) Natural logarithm of permeability relative to its initial value, , with the four displayed states marked. The dimensionless time is , where is the dissolution timescale defined by Budek and Szymczak [64], Eq. (16).
IVP: nonlinear dispersive waves on a periodic network
Model and coupling conditions
We consider nonlinear dispersive wave propagation on a periodic square network comprising vertices and 512 edges of length . Horizontal and vertical edges are oriented to the right and upwards, respectively, with opposite boundaries identified. Each vertex therefore connects two incoming and two outgoing edges.
In the network calculus, denotes the prescribed direction field. For a continuous real scalar field , we define the regular edgewise curvature
so that , without requiring continuity of at vertices. The evolution satisfies
interpreted against continuous scalar test fields. On each edge, this gives the ordinary nonlinear KdV equation,
The nonlinear term produces amplitude-dependent propagation, while the third derivative produces dispersion. At a vertex , label the incoming edges , and the outgoing edges , . Continuity of gives
The vertex contribution to the divergence balances the incoming and outgoing fluxes. Because is continuous and the crossing is balanced, the contributions cancel, leaving
We complete the coupling by prescribing derivative conditions
Together, these are the balanced , conditions considered by Cavalcante and Marques [65 Example 8.3]. For smooth solutions, testing the network balance with unity gives conservation of mass. Testing with and using the slope pairings gives conservation of the squared norm:
We monitor deviations from their initial values to assess numerical conservation.
Pulse evolution
We initialize a localized pulse on one horizontal edge ,
and set elsewhere. We use a Legendre basis with 64 modes per edge. Junction conditions are enforced using correction functions constructed from their adjoints with respect to the physical inner product. Time integration uses the RKGFY scheme with a timestep of , and nonlinear terms are evaluated with 3/2 dealiasing. Resolution is checked by spatial and temporal refinement. Over , the maximum relative drifts in total mass and squared norm are and , respectively.
Figure 5 shows the evolution at . The initially nonnegative pulse develops oscillatory profiles of both signs and transfers amplitude between horizontal and vertical edges. Although the initial disturbance occupies a single horizontal edge, vertical edges carry approximately 52% of the total squared norm by . By , the positive peak has fallen from 1 to approximately 0.157. The conservation measurements show that this decrease accompanies redistribution across the network with negligible loss of total squared amplitude.

Figure 5. KdV pulse evolution on a 16 × 16 periodic network. Gold and blue indicate positive and negative , respectively; pulse width is proportional to . Thin lines show the network geometry. The edge orientation field, , is rightward on horizontal edges and upward on vertical edges. The outgoing endpoints at the right and top boundaries are identified with the corresponding vertices on the left and bottom boundaries.
To distinguish the effect of nonlinearity, we compare with a linear control obtained by omitting , retaining the same initial pulse and coupling conditions. We measure concentration across edges by
This quantity equals one when the squared amplitude is confined to one edge and equals when it is distributed equally among edges. At , is approximately 89 for the nonlinear solution and 203 for the linear control. The nonlinear solution retains a more concentrated distribution across edges at this time.
Discussion
We have developed a concise yet general numerical framework for solving PDEs on physical networks. The contributions include a robust conceptual framework for defining edge differential equations and consistent coupling conditions at vertices, as well as a high-performance distributed spectral solver for simulation.
The network calculus allows differential equations to be written in terms of fields over the whole network, with edge and vertex equations appearing in the same expressions. The practical value is that relations between these contributions remain explicit as a model is developed. Integration by parts relates differential operators to conservation laws, and field evaluations at vertices and their adjoints connect kinematic constraints to reaction forces. The beam example requires a choice of joint kinematics, the KdV example requires directional data and slope pairings, and the mineral model assumes complete mixing in addition to species-flux conservation.
This formulation builds upon Dedalus’s existing field and equation interface. Cardinal bases collect the edge expansions into network fields, while the connectivity maps supply the relations between edge and vertex data. Polynomial differentiation, nonlinear evaluation, time integration and eigenvalue algorithms reuse the existing Dedalus machinery.
The example problems of the paper demonstrate the different capabilities of the infrastructure at a variety of scales. The photonic EVP reproduces the reference spectrum and then computes selected resonances on networks with approximately edges. The mineral example couples pressure and transport LBVP solves to pore evolution, producing channels and increased permeability. The KdV IVP calculation uses the full functionality of the network calculus. From an initial localized pulse, the solution preferentially splits in the northward and eastward directions selected by the edge orientations. Through periodicity of the square grid, backward-propagating dispersive radiation interacts with the positive-amplitude pulse fragments as the solution spreads. The long-time dynamics remain unresolved: the evolution could approach a broadly distributed dispersive state or exhibit pulse recombination and recurrence.
Finally, we envisage ongoing maintenance and improvements as applications require. Adaptive spectral resolution could concentrate computational effort on edges where finer structure develops—this could be particularly useful on hierarchical or multiscale networks—while stronger interface preconditioners could improve parallel scaling. For evolving networks, adapting the distribution of work as the geometry changes may become worthwhile.
Code and data availability
The source code, example scripts, and data needed to reproduce the figures are available at https://github.com/sidneyholden1/dedalus-networks.
Declaration of AI use
The authors conceived the study, devised the overall algorithms, designed the Dedalus computational interfaces, and selected the examples and drafted the manuscript. OpenAI’s ChatGPT and Codex were used to assist with literature searches and identification of references, implementation of numerical routines, collection of simulation data, preparation of figures, and revision of the manuscript. The authors retain full responsibility for the mathematical and computational results, their interpretation, and the content of the manuscript.
References
- [1]K.J. Burns, G.M. Vasil, J.S. Oishi, D. Lecoanet, and B.P. Brown. Dedalus: A flexible framework for numerical simulations with spectral methods. Physical Review Research, 2(2), April 2020.DOI
- [2]G.M. Vasil, K.J. Burns, D. Lecoanet, S. Olver, B.P. Brown, and J.S. Oishi. Tensor calculus in polar coordinates using Jacobi polynomials. Journal of Computational Physics, 325:53–73, November 2016.DOI
- [3]Geoffrey M. Vasil, Daniel Lecoanet, Keaton J. Burns, Jeffrey S. Oishi, and Benjamin P. Brown. Tensor calculus in spherical coordinates using jacobi polynomials. part-i: Mathematical analysis and derivations. Journal of Computational Physics: X, 3:100013, 6 2019.arxiv.org/abs/1804.10320
- [4]Daniel Lecoanet, Geoffrey M. Vasil, Keaton J. Burns, Benjamin P. Brown, and Jeffrey S. Oishi. Tensor calculus in spherical coordinates using jacobi polynomials. part-ii: Implementation and examples. Journal of Computational Physics: X, 3:100012, 6 2019.arxiv.org/abs/1804.09283
- [5]Jeffrey S. Oishi, Benjamin P. Brown, Keaton J. Burns, Daniel Lecoanet, and Geoffrey M. Vasil. Perspectives on reproducibility and sustainability of open-source scientific software from seven years of the dedalus project, 2018.arxiv.org/abs/1801.08200
- [6]The MICrONS Consortium. Functional connectomics spanning multiple areas of mouse visual cortex. Nature, 640:435–447, 2025.DOI
- [7]Xiang Ji, Tiago Ferreira, Beth Friedman, Rui Liu, Hannah Liechty, Erhan Bas, Jayaram Chandrashekar, and David Kleinfeld. Brain microvasculature has a common topology with local differences in geometry that match metabolic load. Neuron, 109(7):1168–1187.e13, 2021.DOI
- [8]A. Cerbelaud, J. Wade, C. H. David, M. Durand, R. P. M. Frasson, T. Pavelsky, and H. Oubanas. Wide-swath altimetry maps bank shapes and storage changes in global rivers. Nature, 651:666–671, 2026.DOI
- [9]Gregory Berkolaiko and Peter Kuchment. Introduction to Quantum Graphs. American Mathematical Society, December 2012.DOI
- [10]Peter Kuchment. Quantum graphs: I. some basic structures. Waves in Random Media, 14(1): S107–S128, January 2004.DOI
- [11]Peter Kuchment. Quantum graphs: An introduction and a brief survey, 2008.DOI
- [12]Lucas Böttcher and Mason A. Porter. Dynamical processes on metric networks. SIAM Journal on Applied Dynamical Systems, 24(4):2848–2885, October 2025.DOI
- [13]Joachim von Below. A characteristic equation associated to an eigenvalue problem on c2-networks. Linear Algebra and its Applications, 71:309–325, November 1985.DOI
- [14]Joachim von Below. Classical solvability of linear parabolic equations on networks. Journal of Differential Equations, 72(2):316–337, April 1988.DOI
- [15]V Kostrykin and R Schrader. Kirchhoff’s rule for quantum wires. Journal of Physics A: Mathematical and General, 32(4):595–630, January 1999.
- [16]Carsten Schubert, Christian Seifert, Jürgen Voigt, and Marcus Waurick. Boundary systems and (skew-)self-adjoint operators on infinite metric graphs. Mathematische Nachrichten, 288 (14-15):1776–1785, May 2015.arxiv.org/abs/1308.2635
- [17]Amru Hussein, David Krejčiřík, and Petr Siegl. Non-self-adjoint graphs. Transactions of the American Mathematical Society, 367(4):2921–2957, August 2014.DOI
- [18]Mark I. Freidlin and Alexander D. Wentzell. Diffusion processes on graphs and the averaging principle. The Annals of Probability, 21(4):2215–2245, October 1993.DOI
- [19]Pavel Exner and Olaf Post. Convergence of spectra of graph-like thin manifolds. Journal of Geometry and Physics, 54(1):77–115, May 2005.DOI
- [20]Marjeta Kramar Fijavž, Delio Mugnolo, and Eszter Sikolya. Variational and semigroup methods for waves and diffusion in networks. Applied Mathematics and Optimization, 55(2):219–240, March 2007.
- [21]Delio Mugnolo. Semigroup Methods for Evolution Equations on Networks. Springer International Publishing, 2014.DOI
- [22]Delio Mugnolo and Silvia Romanelli. Dynamic and generalized wentzell node conditions for network equations. Mathematical Methods in the Applied Sciences, 30(6):681–706, November 2006.DOI
- [23]Marjeta Kramar Fijavž, Delio Mugnolo, and Serge Nicaise. Dynamic transmission conditions for linear hyperbolic systems on networks. Journal of Evolution Equations, 21(3):3639–3673, May 2021.
- [24]Marjeta Kramar Fijavž, Delio Mugnolo, and Serge Nicaise. Linear hyperbolic systems on networks: well-posedness and qualitative properties. ESAIM: Control, Optimisation and Calculus of Variations, 27:7, 2021.DOI
- [25]Jacek Banasiak and Adam Błoch. Telegraph systems on networks and port-hamiltonians. i. boundary conditions and well-posedness. Evolution Equations and Control Theory, 11(4):1331, 2022.arxiv.org/abs/2102.07481
- [26]Federica Gregorio and Delio Mugnolo. Higher-order operators on networks: Hyperbolic and parabolic theory. Integral Equations and Operator Theory, 92(6), November 2020.DOI
- [27]Diego Noja. Nonlinear schrödinger equation on graphs: recent results and open problems. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 372(2007):20130002, January 2014.
- [28]Adilbek Kairzhan, Diego Noja, and Dmitry E Pelinovsky. Standing waves on quantum graphs. Journal of Physics A: Mathematical and Theoretical, 55(24):243001, May 2022.DOI
- [29]Riccardo Adami, Enrico Serra, and Paolo Tilli. Nls ground states on graphs. Calculus of Variations and Partial Differential Equations, 54(1):743–761, December 2014.arxiv.org/abs/1406.4036
- [30]Delio Mugnolo, Diego Noja, and Christian Seifert. Airy-type evolution equations on star graphs. Analysis & amp; PDE, 11:1625–1652, May 2018.
- [31]Márcio Cavalcante. The korteweg–de vries equation on a metric star graph. Zeitschrift für angewandte Mathematik und Physik, 69(5), 9 2018.DOI
- [32]Jaime Angulo and Márcio Cavalcante. Dynamics of the korteweg–de vries equation on a balanced metric graph. Bulletin of the Brazilian Mathematical Society, New Series, 56(1), December 2024.
- [33]Delio Mugnolo and Jean-François Rault. Construction of exact travelling waves for the benjamin-bona-mahony equation on networks. Bulletin of the Belgian Mathematical Society - Simon Stevin, 21(3):415–436, August 2014.DOI
- [34]Roy H. Goodman. Nls bifurcations on the bowtie combinatorial graph and the dumbbell metric graph. Discrete and Continuous Dynamical Systems, 39(4):2203–2232, 2019.arxiv.org/abs/1710.00030
- [35]Adilbek Kairzhan, Robert Marangell, Dmitry E. Pelinovsky, and Ke Liang Xiao. Standing waves on a flower graph. Journal of Differential Equations, 271:719–763, January 2021.arxiv.org/abs/2003.09397
- [36]S.J. Sherwin, V. Franke, J. Peiró, and K. Parker. One-dimensional modelling of a vascular network in space-time variables. Journal of Engineering Mathematics, 47(3-4):217–250, December 2003.DOI
- [37]Mapundi K. Banda, Michael Herty, and Axel and Klar. Gas flow in pipeline networks. Networks & Heterogeneous Media, 1(1):41–56, 2006.DOI
- [38]Michael and Herty. Modeling, simulation and optimization of gas networks with compressors. Networks & Heterogeneous Media, 2(1):81–97, 2007.DOI
- [39]G. M. Coclite, M. Garavello, and B. Piccoli. Traffic flow on a road network. SIAM Journal on Mathematical Analysis, 36(6):1862–1886, January 2005.DOI
- [40]M. Wallace, R. Feres, and G. Yablonsky. Reaction–diffusion on metric graphs: From 3d to 1d. Computers & Mathematics with Applications, 73(9):2035–2052, May 2017.
- [41]Dhruv Saxena, Alexis Arnaudon, Oscar Cipolato, Michele Gaio, Alain Quentel, Sophia Yaliraki, Dario Pisignano, Andrea Camposeo, Mauricio Barahona, and Riccardo Sapienza. Sensitivity and spectral control of network lasers. Nature Communications, 13(1), October 2022.DOI
- [42]Douglas Arnold, Richard Falk, and Ragnar Winther. Finite element exterior calculus: from hodge theory to numerical stability. Bulletin of the American Mathematical Society, 47(2):281–354, January 2010.DOI
- [43]James M. Hyman and Mikhail Shashkov. The orthogonal decomposition theorems for mimetic finite difference methods. SIAM Journal on Numerical Analysis, 36(3):788–818, January 1999.DOI
- [44]Daniele A. Di Pietro, Jérôme Droniou, Kaibo Hu, and Arax Leroy. Design and homological analysis of twisted and bgg stokes–de rham complexes on polygonal meshes, 2025.arxiv.org/abs/2507.17333
- [45]Kaibo Hu, Ting Lin, and Qian Zhang. Distributional hessian and divdiv complexes on triangulation and cohomology. SIAM Journal on Applied Algebra and Geometry, 9(1):108–153, February 2025.arxiv.org/abs/2311.15482
- [46]Daniele Boffi, Kaibo Hu, Yizhou Liang, and Umberto Zerbinati. Finite element exterior calculus for spectra and pseudospectra of advection-diffusion of differential forms, 2026.arxiv.org/abs/2610.01501
- [47]Mario Arioli and Michele Benzi. A finite element method for quantum graphs. IMA Journal of Numerical Analysis, 38(3):1119–1163, 6 2017.DOI
- [48]M. Brio, J.-G. Caputo, and H. Kravitz. Spectral solutions of pdes on networks. Applied Numerical Mathematics, 172:99–117, February 2022.arxiv.org/abs/2104.15048
- [49]Robert Carlson. A quantum graph fit with applications to partial differential equations on networks. Results in Applied Mathematics, 29:100670, February 2026.
- [50]Mihály Kovács and Mihály Vághy. Neumann-neumann type domain decomposition of elliptic problems on metric graphs. BIT Numerical Mathematics, 65(2), May 2025.arxiv.org/abs/2402.05707
- [51]H. Egger, T. Kugler, B. Liljegren-Sailer, N. Marheineke, and V. Mehrmann. On structure-preserving model reduction for damped wave propagation in transport networks. SIAM Journal on Scientific Computing, 40:A331–A365, January 2018.arxiv.org/abs/1704.03206
- [52]David Bolin, Mihály Kovács, Vivek Kumar, and Alexandre Simas. Regularity and numerical approximation of fractional elliptic differential equations on compact metric graphs. Mathematics of Computation, 93:2439–2472, December 2023.arxiv.org/abs/2302.03995
- [53]Christophe Besse, Romain Duboscq, and Stefan Le Coz. Numerical simulations on nonlinear quantum graphs with the grafidi library. The SMAI Journal of computational mathematics, 8:1–47, April 2022.arxiv.org/abs/2103.09650
- [54]Roy H. Goodman, Grace Conte, and Jeremy L. Marzuola. Qglab: A matlab package for computations on quantum graphs. SIAM Journal on Scientific Computing, 47:B428–B453, March 2025.arxiv.org/abs/2401.00561
- [55]Shrirang Abhyankar, Getnet Betrie, Daniel Adrian Maldonado, Lois C. Mcinnes, Barry Smith, and Hong Zhang. Petsc dmnetwork: A library for scalable network pde-based multiphysics simulations. ACM Transactions on Mathematical Software, 46(1):1–24, March 2020.
- [56]Joel Friedman and Jean-Pierre Tillich. Calculus on graphs, 2004.DOI
- [57]P. Kurasov. Distribution theory for discontinuous test functions and differential operators with generalized coefficients. Journal of Mathematical Analysis and Applications, 201(1):297–323, 7 1996.DOI
- [59]Keaton J. Burns, Daniel Fortunato, Keith Julien, and Geoffrey M. Vasil. Corner cases of the tau method: Symmetrically imposing boundary conditions on hypercubes. SIAM Journal on Scientific Computing, 48(4):A2233–A2257, August 2026.arxiv.org/abs/2211.17259
- [60]Keaton J. Burns, Ilaria Fontana, Daniel Lecoanet, and Geoffrey M. Vasil. Characterization of the tau spectrum for eigenvalue problems, 2026.arxiv.org/abs/2610.40307
- [61]Peter J. Olver. Boundary conditions and null lagrangians in the calculus of variations and elasticity. Journal of Elasticity, 155(1-5):75–108, 7 2022.DOI
- [62]Federica Gregorio and Delio Mugnolo. Bi-laplacians on graphs and networks. Journal of Evolution Equations, 20(1):191–232, 7 2019.DOI
- [63]Lars Gottesbüren, Tobias Heuer, Nikolai Maas, Peter Sanders, and Sebastian Schlag. Scalable high-quality hypergraph partitioning. ACM Transactions on Algorithms, 20:1–54, January 2024.arxiv.org/abs/2303.17679
- [64]Agnieszka Budek and Piotr Szymczak. Channeling, clogging and permeability oscillations: Different macroscopic regimes in mineral replacement. Journal of Geophysical Research: Solid Earth, 130(6):e2024JB031081, 2025.
- [65]Márcio Cavalcante and José Marques. The Korteweg–de Vries equation on general star graphs. arXiv:2502.07967v2, 2025.DOI



