{
  "id": "finite-volume-method",
  "title": "Finite-Volume Method",
  "lang": "en",
  "source_lang": "en",
  "summary": "The finite-volume method approximates conservation laws through balances of numerical fluxes across the boundaries of discrete control volumes.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Finite volume method",
    "control-volume method",
    "control-volume discretization"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/finite-volume-method/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/finite-volume-method/",
      "json": "https://aiwiki.page/zh/finite-volume-method.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The finite-volume method is a family of numerical techniques for solving partial differential equations by enforcing balance laws over a collection of small regions called control volumes. Instead of approximating differential operators directly at isolated points, it relates the change in a quantity stored inside each volume to fluxes through its boundary and sources within it. Its defining feature is local conservation: the numerical amount leaving one control volume through a shared face is exactly the amount entering its neighbor.\n\nFinite-volume methods are central to computational fluid dynamics, where conservation of mass, momentum, and energy must remain meaningful even across shocks and material interfaces. They are also used for heat transfer, groundwater flow, chemical transport, semiconductor modeling, and geophysical simulation. The approach accommodates structured and unstructured meshes, stationary and moving grids, and equations ranging from linear diffusion to nonlinear systems of conservation laws. Its practical behavior depends on how face fluxes, spatial reconstruction, time integration, and boundary conditions are combined."
    },
    {
      "heading": "Integral conservation and the discrete balance",
      "text": "Consider a conservation law for a scalar quantity or a vector of conserved quantities \\(U\\):\n\n\\[\\frac{\\partial U}{\\partial t}+\\nabla\\cdot F(U)=S.\\]\n\nHere \\(F\\) is the physical flux and \\(S\\) is a source term. For fluid dynamics, the components of \\(U\\) may represent mass density, momentum density, and total energy density. Integrating over a fixed control volume \\(K\\), and applying the divergence theorem, gives\n\n\\[\\frac{d}{dt}\\int_K U\\,dV + \\int_{\\partial K}F(U)\\cdot n\\,dA = \\int_K S\\,dV,\\]\n\nwhere \\(n\\) is the outward unit normal. This integral statement is the starting point of the discretization. It remains meaningful in circumstances where the solution is not differentiable, provided the fluxes and integrals are appropriately defined.\n\nThe primary unknown in a cell-centered scheme is usually the cell average,\n\n\\[\\overline U_K=\\frac{1}{|K|}\\int_K U\\,dV.\\]\n\nFor a control volume with planar faces \\(f\\), a typical semidiscrete approximation is\n\n\\[|K|\\frac{d\\overline U_K}{dt} + \\sum_{f\\subset\\partial K}|f|\\,\\widehat F_{K,f} = |K|\\,\\overline S_K.\\]\n\nThe numerical flux \\(\\widehat F_{K,f}\\) approximates the physical flux in the outward normal direction, averaged over a face. More elaborate formulas use several quadrature points on each face, but the underlying balance is unchanged. The word semidiscrete indicates that space has been discretized while time remains continuous.\n\nFor neighboring cells \\(K\\) and \\(L\\), conservation requires their shared-face contributions to be equal and opposite:\n\n\\[\\widehat F_{K,f}=-\\widehat F_{L,f}.\\]\n\nConsequently, summing the cell equations cancels all internal fluxes. Only exterior-boundary fluxes and volume sources affect the total amount in the computational domain. In software this cancellation is commonly secured by computing each internal face flux once and adding it to the two neighboring residuals with opposite signs.\n\nThis algebraic property is distinct from accuracy. A very diffusive approximation can be exactly conservative, while an accurate pointwise approximation may fail to conserve the desired quantity. It is also important to distinguish a cell average from a value at the cell center. For a smooth field on a regular cell, the two may differ only by a second-order term, but that difference matters in high-order schemes. Finite-volume reconstruction must reproduce averages, rather than silently reinterpret them as point samples."
    },
    {
      "heading": "Control volumes, mesh geometry, and data placement",
      "text": "A finite-volume mesh partitions the computational domain into cells whose interiors do not overlap. In two dimensions these may be triangles or general polygons; in three dimensions they may be tetrahedra, hexahedra, prisms, or other polyhedra. The basic balance needs cell volumes, face areas, face normals, and connectivity between neighboring cells. Reconstruction and diffusive flux calculations additionally require geometric information such as centroids, distances, and interpolation weights.\n\nStructured meshes possess a regular logical indexing, even when their physical cells are curved or stretched. Their predictable connectivity supports compact storage and efficient access to neighboring values. Unstructured meshes describe connectivity explicitly and can fit complicated boundaries or concentrate cells near selected features. Neither arrangement alone determines accuracy: a highly distorted structured grid can be more difficult numerically than a well-shaped unstructured grid.\n\nIn a cell-centered formulation, the principal unknowns are associated with the primal mesh cells. A vertex-centered formulation instead constructs control volumes around mesh vertices, often by joining cell centers, face centers, and edge midpoints. The resulting dual mesh supplies the integration regions. Staggered arrangements place different variables at different geometric locations—for example, pressure in cells and normal velocity components on faces. Such arrangements can help enforce the coupling between pressure and velocity in incompressible flow.\n\nGeometric identities have direct numerical consequences. For a closed polyhedral cell,\n\n\\[\\sum_{f\\subset\\partial K}|f|\\,n_{K,f}=0.\\]\n\nThus, a spatially constant physical flux has zero net contribution when the geometry is represented consistently. Violating this identity through inconsistent geometric approximations can generate artificial residuals even for a uniform state. Curved boundaries and high-order face geometry therefore require suitable integration, rather than merely higher-order reconstruction of the solution.\n\nMesh quality affects both accuracy and robustness. Strong skewness complicates interpolation from cell centers to faces, while nonorthogonality complicates approximation of normal derivatives. Very small cells can impose severe explicit time-step restrictions. Anisotropic cells may nevertheless be appropriate: thin, stretched cells are useful for resolving boundary layers, provided the discretization accounts for their orientation and aspect ratio.\n\nIn adaptive mesh refinement, selected regions are refined as shocks, interfaces, or other structures evolve. At an interface between coarse and fine cells, conservation requires the integrated coarse-face flux to equal the sum of the corresponding fine-face fluxes. Conservative restriction, prolongation, and flux correction are therefore essential parts of refinement algorithms, not optional bookkeeping."
    },
    {
      "heading": "Numerical fluxes for transport and nonlinear waves",
      "text": "For transport-dominated equations, the main spatial problem is determining what crosses a face when neighboring cells provide different states. In one-dimensional linear advection,\n\n\\[u_t+a u_x=0,\\]\n\nthe physical direction of propagation provides a natural answer. The first-order upwind flux is\n\n\\[\\widehat f(u_L,u_R)= \\begin{cases} a u_L,&a\\geq 0,\\\\ a u_R,&a<0. \\end{cases}\\]\n\nInformation comes from the left when \\(a>0\\) and from the right when \\(a<0\\). A centered flux averages both sides, but without suitable dissipation or an appropriate accompanying discretization it may produce unstable or oscillatory approximations.\n\nFor nonlinear conservation laws, the propagation direction depends on the state. A Riemann problem consists of a conservation law with two constant initial states separated by a discontinuity. Its solution may contain shocks, rarefactions, and contact discontinuities. Godunov-type finite-volume methods use exact or approximate Riemann solutions at cell interfaces to construct a flux that reflects these wave motions. On a multidimensional mesh, the local problem is generally posed in the direction normal to a face.\n\nA widely used approximate flux is the local Lax–Friedrichs, or Rusanov, flux:\n\n\\[\\widehat f(u_L,u_R) = \\frac{f(u_L)+f(u_R)}{2} - \\frac{\\alpha}{2}(u_R-u_L),\\]\n\nwhere \\(\\alpha\\) bounds the magnitudes of relevant characteristic speeds. For systems, these speeds are associated with the eigenvalues of the normal flux Jacobian matrix. The second term supplies numerical dissipation. Overestimating \\(\\alpha\\) generally makes the method more diffusive; an inadequate bound can undermine stability.\n\nOther approximate solvers resolve more of the wave structure. Roe’s solver linearizes the system at a specially chosen averaged state. Harten–Lax–van Leer solvers approximate the wave fan through a small number of representative waves; HLLC restores an intermediate contact wave for suitable systems, notably the compressible Euler equations. These choices affect contact resolution, robustness near vacuum, and sensitivity to strong grid-aligned shocks. No single flux is uniformly best for every regime.\n\nConsistency requires \\(\\widehat f(u,u)=f(u)\\). Conservation requires compatible fluxes on either side of every shared interface. Neither condition alone ensures the physically relevant solution. Nonlinear conservation laws can admit multiple weak solutions, so an additional entropy admissibility condition is needed. A numerical flux must therefore be judged not only by its formal accuracy, but also by its treatment of shocks and its compatibility with the physical admissibility conditions of the governing system."
    },
    {
      "heading": "Reconstruction, limiters, and high-order accuracy",
      "text": "A piecewise-constant representation provides only first-order spatial accuracy in the usual Godunov construction. Higher-order finite-volume methods reconstruct a spatial profile inside each cell from neighboring averages. A linear reconstruction has the form\n\n\\[U_K(x)=\\overline U_K+g_K\\cdot(x-x_K),\\]\n\nwhere \\(x_K\\) is the cell centroid and \\(g_K\\) is an estimated gradient. Choosing the centroid ensures that the linear correction has zero cell average. Reconstructed values on opposite sides of a face become the left and right inputs to the numerical flux.\n\nGradients may be estimated by Green–Gauss formulas or weighted least squares. On an unstructured mesh, a least-squares reconstruction fits differences between neighboring averages to differences between cell positions. The quality of this fit depends on the geometry and distribution of neighboring cells. A stencil that poorly samples one spatial direction may yield an inaccurate or ill-conditioned reconstruction even if it contains many neighbors.\n\nUnrestricted high-order reconstruction tends to oscillate near discontinuities. Such oscillations can create negative density, negative pressure, or concentrations outside their admissible ranges. Slope limiters reduce reconstructed gradients where neighboring averages suggest a lack of smoothness. The minmod limiter, for example, returns the smallest magnitude among slopes of a common sign and returns zero when their signs disagree. More compressive limiters preserve sharper fronts but may be less suitable for smooth extrema or complicated multidimensional structures.\n\nThe total-variation-diminishing principle measures oscillation in one dimension through\n\n\\[TV(u)=\\sum_i |u_{i+1}-u_i|.\\]\n\nUnder suitable assumptions and time-step restrictions, a TVD method does not increase this quantity. It offers a useful route to nonoscillatory transport schemes, but imposes accuracy limitations near smooth extrema. Moreover, a one-dimensional limiter applied separately along coordinate directions does not automatically provide the same guarantees on a general multidimensional mesh.\n\nHigher-order alternatives include piecewise-parabolic reconstruction, essentially nonoscillatory reconstruction, and weighted essentially nonoscillatory reconstruction. WENO methods combine candidate polynomial reconstructions with nonlinear weights based on smoothness indicators. In smooth regions the weights approach values designed for high-order accuracy; near discontinuities they reduce the contribution from unsuitable stencils. Accuracy then depends on the reconstruction, quadrature, numerical flux, and time integrator together.\n\nFor systems, reconstruction can be performed in conserved variables, primitive variables, or local characteristic variables. Characteristic reconstruction often separates interacting wave families more effectively, but requires additional matrix operations and a reliable eigenstructure. Positivity-preserving procedures may further rescale a high-order profile toward its cell average. Such modifications preserve the average while constraining reconstructed states, illustrating how conservation and nonlinear admissibility can be built into the same reconstruction."
    },
    {
      "heading": "Diffusion, elliptic equations, and discontinuous materials",
      "text": "Finite-volume methods are not limited to propagating waves. For diffusion, the transported quantity moves down a gradient. A representative equation is\n\n\\[\\frac{\\partial u}{\\partial t} - \\nabla\\cdot(\\kappa\\nabla u)=s,\\]\n\nwhich includes the heat equation when the coefficients have the appropriate physical interpretation. The outward diffusive flux is \\(-\\kappa\\nabla u\\cdot n\\), so the central numerical task is approximating a normal derivative at each face.\n\nOn an orthogonal mesh with scalar diffusivity, a two-point flux approximation uses the values in the two adjacent cells:\n\n\\[Q_{K,L}=T_{K,L}(u_K-u_L).\\]\n\nHere \\(Q_{K,L}\\) is the integrated outward flux from \\(K\\) to \\(L\\), and \\(T_{K,L}\\) is a transmissibility. If the distance from each cell center to the shared face is \\(d_K\\) or \\(d_L\\), and the conductivity is piecewise constant, a standard expression is\n\n\\[T_{K,L} = \\frac{|f|}{d_K/\\kappa_K+d_L/\\kappa_L}.\\]\n\nThe denominator adds the two normal resistances. This explains why a distance-weighted harmonic average, rather than an arithmetic average, is appropriate for a layered medium with a discontinuous coefficient.\n\nThe two-point formula is not universally consistent. On a nonorthogonal mesh, the vector joining cell centers is not generally parallel to the face normal. With tensor-valued diffusivity, even geometric orthogonality may be insufficient: the relevant alignment depends on the diffusion tensor. Ignoring these distinctions can produce an error that does not disappear at the expected rate under refinement.\n\nSeveral extensions address this problem. Nonorthogonal correction methods split the flux into a two-point part and a correction estimated from reconstructed gradients. Multipoint flux approximations use a larger neighborhood to account for directional coupling. Mixed and mimetic formulations introduce additional flux information or discrete operators designed to preserve structural identities. These alternatives differ in stencil size, coefficient restrictions, monotonicity, and computational cost.\n\nFor steady diffusion, the cell balances produce a system of linear equations. Under suitable geometric and coefficient assumptions, its matrix has a positive diagonal and nonpositive off-diagonal entries, supporting a discrete maximum principle. Strong anisotropy, distorted cells, or certain higher-order corrections can destroy that property, allowing nonphysical undershoots and overshoots even when the equations remain conservative.\n\nMaterial interfaces require physical interface conditions as well as coefficient averaging. With no surface source, the normal flux is continuous across the interface; the transported field may be continuous or may obey a jump law associated with contact resistance or another interface model. A conservative discretization must represent these conditions explicitly rather than treating every coefficient discontinuity as an ordinary smooth region."
    },
    {
      "heading": "Time integration, boundary conditions, and source terms",
      "text": "After spatial discretization, the cell averages satisfy a system of ordinary differential equations, often written\n\n\\[M\\frac{dU}{dt}=R(U,t),\\]\n\nwhere \\(M\\) contains cell volumes and \\(R\\) contains numerical fluxes and sources. This separation permits different time integrators to be paired with the same spatial operator. Forward Euler is inexpensive but first-order accurate. Multistage Runge–Kutta methods provide higher-order explicit integration; strong-stability-preserving variants retain certain convex stability properties of forward Euler under an appropriate step restriction.\n\nExplicit transport schemes are limited by the Courant–Friedrichs–Lewy condition. For one-dimensional advection, the relevant nondimensional quantity is\n\n\\[C=\\frac{|a|\\Delta t}{\\Delta x}.\\]\n\nFirst-order upwind with forward Euler is stable for \\(C\\leq 1\\). On general meshes, local restrictions involve cell volume, face areas, and characteristic speeds. Explicit diffusion introduces a more severe scaling, typically \\(\\Delta t=O(h^2)\\), compared with the \\(O(h)\\) transport restriction for comparable uniform grids.\n\nImplicit methods evaluate some or all residual terms at the new time level. They can remove particular explicit stability restrictions, but each step requires a linear or nonlinear solve, and large steps can still be inaccurate. Implicit–explicit methods treat stiff diffusion or reactions implicitly while advancing selected transport terms explicitly. In chemically reacting flows, the source-term time scales may be far shorter than those associated with fluid motion.\n\nBoundary conditions enter through boundary-face fluxes, reconstructed boundary states, or ghost cells. A prescribed diffusive flux can be inserted directly into the balance; a prescribed value generally requires an approximation linking that value to the adjacent interior state. For hyperbolic systems, the number of independent boundary data is determined by incoming characteristics. Imposing every variable at a subsonic outflow can overconstrain the physical problem.\n\nSource terms need equally careful discretization. In shallow-water flow over uneven terrain, the pressure-related flux gradient must balance the bed-slope source for a stationary lake. A well-balanced scheme preserves such equilibria exactly or to a deliberately controlled accuracy. Without this balance, truncation errors can create spurious motion larger than the small physical disturbance of interest.\n\nMoving control volumes introduce another balance requirement. Their fluxes must account for boundary motion, and their evolving volumes must satisfy a discrete geometric conservation law. Otherwise, mesh motion alone can disturb a uniform flow. Boundary treatment, source integration, and moving geometry are therefore integral parts of the method’s consistency and conservation, not merely peripheral implementation choices."
    },
    {
      "heading": "Stability, convergence, and physical admissibility",
      "text": "The theoretical assessment of a finite-volume method separates several questions. Consistency asks whether the discrete flux and source approximate their continuous counterparts. Stability asks whether perturbations remain controlled in a specified sense. Convergence asks whether the discrete solutions approach a solution of the governing problem as the mesh and time step are refined. Local conservation supports these properties but does not establish them by itself.\n\nFor nonlinear hyperbolic equations, the Lax–Wendroff theorem gives a foundational connection: under suitable assumptions, a convergent sequence generated by a consistent conservative approximation has a weak solution as its limit. The theorem does not prove that the numerical sequence converges, and weak solvability alone does not select the entropy-admissible solution. Additional estimates, compactness arguments, or entropy inequalities are needed, with substantially different difficulties for scalar equations and systems.\n\nA discrete entropy inequality typically takes the form\n\n\\[|K|\\frac{d\\eta(U_K)}{dt} + \\sum_f |f|\\,\\widehat q_{K,f} \\leq 0\\]\n\nfor a source-free problem, where \\(\\eta\\) is a convex mathematical entropy and \\(\\widehat q\\) is a compatible entropy flux. Appropriate boundary terms must also be considered. Entropy-stable fluxes are designed to satisfy such a relation, while entropy-conservative fluxes reproduce an equality before additional dissipation is introduced. Mathematical entropy need not use the same sign convention as thermodynamic entropy.\n\nPositivity and boundedness are related but separate requirements. A compressible-flow method may need to preserve positive density and internal energy; a transport method may need to preserve nonnegative species concentrations. These properties often depend jointly on the flux, reconstruction, and time-step restriction. An entropy-stable method is not automatically positivity preserving, and a positivity-preserving method is not automatically accurate near a smooth extremum.\n\nFormal order should also be distinguished from observed error. A second-order method can display second-order convergence for smooth solutions while converging more slowly in a norm sensitive to shocks. Limiting, under-resolved layers, boundary approximations, and mesh irregularity can all reduce measured rates. Consequently, verification commonly combines manufactured smooth solutions, refinement studies, conservation checks, and discontinuous benchmark problems.\n\nNumerical stability is especially delicate for coupled physical systems. At low Mach number, a compressible flux may introduce excessive dissipation unless it is designed or modified for that limit. Magnetohydrodynamics additionally requires control of the divergence of the magnetic field, while multiphase systems may impose compatibility conditions between pressure, volume fractions, and material interfaces. Preserving one structural property cannot substitute for checking the others."
    },
    {
      "heading": "Algebraic solution, software organization, and applications",
      "text": "A practical finite-volume code often organizes its work around face-based residual assembly. For each internal face, it obtains the neighboring states, reconstructs face values, evaluates the numerical flux, and accumulates equal and opposite contributions. Boundary faces invoke the relevant physical condition. This organization makes conservation explicit and separates mesh connectivity from many details of the governing equations.\n\nImplicit time stepping and steady-state calculations lead to large sparse algebraic systems. A nonlinear residual equation \\(R(U)=0\\) can be attacked with Newton’s method, approximate Newton methods, or segregated iterations. Each nonlinear iteration may require solving a system involving a sparse matrix or applying its action without explicitly storing it. Preconditioning and multigrid methods are often decisive because fine meshes contain error components on widely different spatial scales.\n\nIn incompressible flow, pressure enforces the divergence constraint rather than evolving through an independent compressibility law. Pressure-correction methods derive a pressure or pressure-increment equation from discrete momentum and continuity balances. On collocated meshes, careless interpolation can permit alternating pressure patterns that couple weakly to velocity. Momentum-based face interpolation, associated with the Rhie–Chow approach, suppresses this decoupling while constructing mass fluxes consistent with the momentum discretization.\n\nFinite-volume algorithms suit parallel computing because most fluxes depend on local neighborhoods. A partitioned mesh exchanges neighboring-cell data through halo regions. Communication costs grow with partition boundaries, while implicit solvers introduce additional global coordination. High-order reconstruction enlarges stencils, and adaptive meshes complicate load balancing. On accelerators, data layout, face ordering, memory access, and residual-update conflicts can matter as much as the arithmetic cost of a flux formula.\n\nApplications emphasize different aspects of the same framework. Compressible aerodynamics requires reliable shock capturing and wall treatment. Environmental hydraulics benefits from conservation, wetting-and-drying treatment, and well-balanced shallow-water schemes. Groundwater simulation requires transmissibilities that represent strongly varying geological properties and material interfaces. Heat-transfer calculations may couple convection, conduction, radiation models, and solid–fluid interfaces.\n\nFinite-volume methods are also extensively used for the Navier–Stokes equations with turbulence models, combustion, and multiphase-flow closures. A numerically conservative solution is not necessarily a physically complete prediction: uncertainty may originate in constitutive laws, unresolved turbulence, boundary data, or insufficient spatial resolution. Practical credibility therefore depends on separating discretization verification from validation of the physical model, while monitoring conservation errors, nonlinear convergence, and mesh sensitivity."
    },
    {
      "heading": "Historical development and relations to other methods",
      "text": "Finite-volume ideas developed from engineering control-volume balances and several strands of numerical conservation-law research, rather than from a single universally agreed invention. Early fluid calculations often combined staggered variable arrangements, conservative difference formulas, and cell-based balances. The later distinction between finite differences and finite volumes became clearer as methods were extended to irregular meshes and discontinuous solutions.\n\nSergei Godunov’s 1959 paper introduced a conservative method based on exact local Riemann problems for gas dynamics. Its first-order formulation became a foundation for modern shock-capturing methods. The limitation that a linear monotone scheme cannot be more than first-order accurate helped motivate nonlinear reconstruction and limiting. Higher accuracy could not be obtained simply by extending a fixed linear stencil while retaining all the desired monotonicity properties.\n\nDuring the 1970s and 1980s, separate but interacting developments broadened the method’s reach. Patankar and Spalding introduced the SIMPLE pressure–velocity coupling procedure in 1972, and Suhas Patankar’s 1980 book Numerical Heat Transfer and Fluid Flow helped systematize control-volume methods for engineering applications. Bram van Leer’s 1979 MUSCL formulation provided a major route to higher-resolution Godunov methods. Philip Roe’s approximate Riemann solver appeared in 1981, and Ami Harten’s 1983 work established influential high-resolution and TVD constructions.\n\nSubsequent advances included essentially nonoscillatory methods, the WENO formulation introduced by Xu-Dong Liu, Stanley Osher, and Tony Chan in 1994, and the efficient WENO developments of Guang-Shan Jiang and Chi-Wang Shu in 1996. Alongside shock-capturing research, work on unstructured meshes, multipoint diffusion fluxes, adaptive refinement, entropy stability, and compatible discrete operators extended finite-volume techniques to increasingly complex geometries and physical systems.\n\nThe relationship with the finite-difference method is close but not identical. On a uniform one-dimensional grid, a finite-volume update can have the same algebraic appearance as a conservative finite-difference formula. The distinction lies in the interpretation of the unknowns and operators: averages and integrated balances in one case, point values and derivative approximations in the other. For irregular cells, discontinuities, and high-order reconstruction, that distinction becomes consequential.\n\nThe finite-element method begins from a different framework of trial functions, test functions, and weak formulations, although considerable overlap exists. A lowest-order discontinuous Galerkin method using piecewise-constant functions and compatible numerical fluxes reduces to a basic finite-volume scheme. Higher-order discontinuous Galerkin methods retain additional degrees of freedom within each cell, whereas classical high-order finite-volume methods reconstruct that information from neighboring averages. Hybrid and control-volume finite-element formulations combine elements of both approaches, so the boundaries between method families are defined by their mathematical construction rather than by mesh shape alone."
    }
  ],
  "links": [
    "partial-differential-equation",
    "divergence-theorem",
    "eigenvalues-and-eigenvectors",
    "jacobian-matrix",
    "weak-solution",
    "gradient",
    "least-squares",
    "diffusion",
    "heat-equation",
    "system-of-linear-equations",
    "ordinary-differential-equation",
    "courant-friedrichs-lewy-condition",
    "numerical-stability",
    "magnetic-field",
    "newtons-method",
    "sparse-matrix",
    "multigrid-method",
    "parallel-computing",
    "groundwater",
    "navier-stokes-equations",
    "finite-difference-method",
    "finite-element-method"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}