ASF

RVE-to-Mesh Pipeline for Heterogeneous Continua

April 15, 2026|
A
A

Two earlier posts on this site developed complementary halves of a single picture. Reservoir Geometry took an anisotropic permeability field as given and built a Riemannian manifold on top of it with . Numerical Analysis via Discrete Exterior Calculus took a manifold as given and discretised its Laplace-Beltrami operator using the Hodge star on a simplicial mesh. This post closes the loop on both ends. Upstream of the reservoir post comes from mean-field homogenisation on a representative volume element. Downstream of the DEC post the mesh adapts to the resulting geometry by curvature-driven remeshing with the closed-form scalar curvature .

The result is a single arc from grain-scale microstructure to a converged finite-dimensional solution: microstructure feeds homogenisation, homogenisation produces a tensor field, the tensor field defines a metric, a conformal rescaling of that metric induces the Hodge star that discretises Darcy's operator, and the curvature of the metric tells the mesh where to refine. Every arrow in that chain is a mathematical operation with a closed-form or algorithmic expression and is implemented in, or scheduled for, the cartan workspace.

1. Scale separation and the representative volume element

A porous or fractured rock is heterogeneous on the scale of pores and grains, with length scale on the order of micrometres, and it is homogeneous on the scale of reservoirs, with length scale on the order of hundreds of metres. A representative volume element (RVE) is a sub-volume of the rock at an intermediate scale , satisfying

The RVE is large enough that the statistics of its microstructure (pore sizes, fracture orientations, grain packing) match those of the bulk rock, and small enough that macroscopic fields (pressure, stress) are effectively constant across it. The existence of such an intermediate scale is the scale separation hypothesis. It is the assumption that licenses every homogenisation scheme discussed below.

Definition (Representative volume element).

A sub-volume of a heterogeneous medium is a representative volume element if (i) the microstructural statistics inside are indistinguishable from those of to within a prescribed tolerance, and (ii) the effective properties computed from are independent of boundary conditions imposed on to within the same tolerance.

Introduce the small parameter and the fast variable . A heterogeneous tensor field oscillating on the RVE scale has the form , where is -periodic in . The goal of homogenisation is to produce a slowly varying effective tensor that captures the macroscopic behaviour of as .

2. Two-scale asymptotic homogenisation

The rigorous route to is two-scale asymptotic expansion, developed for periodic structures by Bensoussan, Lions and Papanicolaou [5] and placed on the footing of two-scale convergence, introduced by Nguetseng [11] and developed by Allaire [6]. Consider the Darcy pressure equation on the reservoir and expand the solution in powers of :

with each assumed -periodic in . The spatial derivative splits according to the chain rule: , acting on functions of both slow and fast variables. Substituting into the Darcy equation and collecting powers of produces a hierarchy of problems.

The balance forces to be independent of the fast variable, . The balance gives the cell problem: find , periodic in , such that

The cell problem is a linear elliptic boundary-value problem on the periodic unit cell , parameterised by the slow variable and solved once per macroscopic point (or once per RVE if the microstructure is statistically homogeneous). Its solution is the corrector that encodes the fine-scale response to a unit macroscopic gradient in direction .

The balance yields the homogenised macroscopic equation , with

The cell average is taken with respect to the fast variable; the slow variable is a parameter. Symmetry and positive definiteness of follow from symmetry and positive definiteness of the microstructural tensor . Thus at every .

Remark (Variational form of the cell problem).

The cell problem [eq:cell-problem] is the Euler-Lagrange equation of the variational problem

where is the Sobolev space of -periodic functions. This places in the framework of -convergence: it is the variational limit of the energy functionals associated with .

3. Mean-field homogenisation and the Eshelby localisation

Two-scale asymptotics requires solving the cell problem numerically on each RVE. For fractured and porous media with a small number of inclusion families (matrix, pores, cracks with a few orientation classes), a closed-form alternative is available: mean-field homogenisation in the Zaoui [2] and Ponte Castañeda-Willis [4] tradition. The starting point is the Eshelby inclusion problem [3].

Theorem (Eshelby (1957)).

Let be an ellipsoidal inclusion (characterised by its shape tensor ) embedded in an infinite homogeneous matrix with conductivity tensor . If a uniform source is applied inside the inclusion, the induced gradient field inside the inclusion is uniform:

The second-order tensor , called the Hill polarisation tensor, depends on the inclusion shape and the matrix conductivity but not on the inclusion itself.

For a sphere in an isotropic matrix (, ), the Hill tensor is . Closed forms for general spheroids in isotropic and transversely isotropic matrices are tabulated in the ECHOES documentation [10] and in Milton's monograph [1].

From Eshelby's theorem one constructs, for each inclusion family with volume fraction and conductivity contrast , a dilute localisation tensor

The effective conductivity is then a weighted combination of the . Three schemes cover most use cases.

Dilute scheme. Non-interacting inclusions in a host matrix:

Mori-Tanaka scheme. Each inclusion sits in the matrix and sees the average field of the matrix phase. Concentration tensors then sum to the identity:

Self-consistent scheme. Each phase is embedded in the effective medium itself. This gives an implicit equation solved by fixed-point iteration:

Each scheme produces an SPD tensor field at every macroscopic point. For a fractured reservoir whose crack statistics (density, orientation distribution, aspect ratio) vary with Mori-Tanaka gives the effective permeability field the reservoir post takes as input.

Remark (Two routes to one output).

Two-scale asymptotics and mean-field homogenisation both produce . Asymptotics is rigorous (a theorem of -convergence) but needs numerical cell-problem solutions. Mean-field is closed-form but leans on the Eshelby geometry (ellipsoidal inclusions, infinite matrix). In practice asymptotics validates mean-field on canonical microstructures, and mean-field is used when inclusion statistics are all that is measured. The downstream geometry is independent of the route that produced .

4. Metric and its fibre bundle

Given at every , define the Riemannian metric

This is the construction developed in the reservoir geometry post. The reservoir becomes a Riemannian manifold . Its geodesics are the rays of the pressure front, and permeability contrasts refract them. Darcy streamlines are the integral curves of , the gradient lines of the pressure in the metric , and they are geodesics only in special cases such as uniform flow in a homogeneous layer.

Proposition (Darcy's operator as a Laplace-Beltrami operator).

Let be a smooth -valued field on and Darcy's operator.

(i) For , . The two operators agree when is constant.

(ii) For and , for every . Equivalently .

(iii) For no metric of the form with has proportional to unless is constant.

Proof

For with a function, and . Hence

With the Laplace-Beltrami formula gives , and the product rule gives (i). For the choice makes [eq:conformal-weight] equal to and . This is (ii). The 3-form identity is the coordinate expression . For the exponent vanishes and [eq:conformal-weight] is for every . Then . The first-order term vanishes for all only if . This is (iii). □

Part (i) is the correction formula of §4.2 of the reservoir post, where the drift term is the reason the Darcy divergence is the physical operator for a heterogeneous formation. Part (ii) removes the drift in three dimensions. The pipeline therefore keeps two metrics in one conformal class. The metric defines the geometry of the reservoir post (distances, geodesics and the curvature used for refinement in section 6), and supplies the Hodge star that discretises the pressure equation in section 5. The identities in the proof were checked symbolically with SymPy for diagonal in 3D and for a general symmetric matrix.

The structure is cleaner when viewed as a fibre bundle. Let be the bundle whose fibre at is , the space of symmetric positive definite matrices. The effective permeability is a section of this bundle,

Two operations sit on this bundle.

Base-horizontal operations are calculus along with the metric at each point: the Levi-Civita connection, the Laplace-Beltrami operator, geodesics, gradient streamlines, scalar curvature. These are the content of the reservoir post.

Fibre-vertical operations are geometry inside at a fixed base point: the affine-invariant Riemannian metric on [9], geodesic interpolation between measured values at neighbouring wells, the Fréchet mean of a distribution of SPD samples. These are used when is known only at discrete measurement points and must be interpolated to the rest of .

Definition (Affine-invariant metric on ).

For tangent vectors (symmetric matrices) at , the affine-invariant inner product is

The geodesic from to is

and the geodesic distance is .

The same structure that parameterises the fibre is used by every stage of the pipeline: homogenisation outputs a point in per RVE, the metric inverts that point at each , the Hodge star (next section) is parameterised by the same tensor through , and mesh quality at each triangle is measured against the same inner product.

5. Hodge star from the homogenised metric

The DEC post developed discrete exterior calculus on a flat Euclidean background. It follows the framework of Hirani [7] and Desbrun, Hirani, Leok and Marsden [8]. The extension to a Riemannian manifold is where the metric enters the discretisation. By Proposition [darcy-metric] the metric to use for the pressure equation is , and in the rest of this section stands for any Riemannian metric with as the case of interest. In the continuous setting the Hodge star is defined by

where the pointwise inner product on -forms is constructed from and its inverse (for , ), and is the Riemannian volume form. The Laplace-Beltrami operator on 0-forms is . For the 3-form is Darcy's operator times the Euclidean volume form. The discrete pressure equation is therefore a stiffness matrix built from of with a load vector integrated over Euclidean dual cells.

On a simplicial mesh of the Hodge star becomes a diagonal (or block-diagonal) linear map between cochain spaces. For the scalar Poisson problem on vertices the diagonal circumcentric Hodge star maps a primal edge of length (measured in the metric ) to its dual edge of length , with ratio

Every geometric quantity in [eq:hodge-diagonal] is computed in the metric induced by . The length of a primal edge from vertex to vertex is

with a geodesic in . For a fine mesh the straight-line approximation

with the midpoint is second-order accurate in the mesh size and matches the order of the linear finite-element discretisation used downstream.

Remark (Anisotropic Hodge stars).

Freeze at a constant value on each simplex. The linear map is an isometry from that simplex with metric onto a Euclidean simplex. Circumcentres, dual volumes and the diagonal star of [eq:hodge-diagonal] are those of the image. Off-diagonal entries of are no obstacle by themselves. The obstacle is the sign of the diagonal entries. A mesh that is well-centred in the Euclidean sense need not be well-centred after the stretch by , and on a tetrahedron the unique consistent diagonal can have negative entries. It has a negative entry on each of the six tetrahedra of the standard subdivision of a cube about its main diagonal. The remedy is a non-diagonal star. The Galerkin star on Whitney 1-forms in the metric gives equal to the linear finite-element stiffness matrix for piecewise-constant , which is symmetric positive semidefinite on any mesh. In the cartan workspace cartan-mimetic builds consistent positive definite local stars on Euclidean simplices of any dimension and reports whether the diagonal star is usable on a given simplex, and cartan-matfree applies Galerkin mass operators for a piecewise-constant metric given by squared edge lengths. cartan-dec's circumcentric star covers triangle meshes only.

The stiffness matrix assembled with of is the discrete Darcy operator. Its null space consists of constants (for Neumann boundary conditions), its spectrum depends on , and its condition number grows with the anisotropy ratio of . Every property of the Darcy operator that matters for pressure diffusion enters through the Hodge star.

6. Scalar curvature as an a priori refinement indicator

The closed-form scalar curvature of the isotropic permeability metric takes the homogenisation output to the refinement indicator used by cartan-remesh.

Theorem (Scalar curvature of the homogenised permeability metric).

Let be a smooth positive scalar field, and let be the conformally flat Riemannian metric on with components in Cartesian coordinates. Then the scalar curvature of is

where and are the flat Laplacian and gradient on .

Proof

The metric is conformally flat: with conformal factor

Its inverse is and its determinant is .

Christoffel symbols. The Levi-Civita connection of a conformally flat metric has

This follows from the general definition by direct substitution of and cancellation.

Conformal transformation of scalar curvature. Under a conformal change in dimension , the scalar curvatures are related by the classical formula

derived in Besse, Einstein Manifolds, §1.J. Specialising to flat background , and :

Expression in . From ,

For the Laplacian, compute

Assembly. Substitute into [eq:R-in-phi]:

Distribute:

Collecting the last two terms gives the coefficient :

which is [eq:R-scalar]. □

The formula [eq:R-scalar] has inputs produced by homogenisation (section 3). Its output drives the refinement criterion below. The remesher needs no more than two derivatives of the homogenised permeability.

Regions of high are regions where the metric varies rapidly. There geodesics bend, Jacobi fields spread or focus and pressure transients accumulate error. The indicator is a heuristic: measures the variation of and enters no error estimate for the discrete solution. A finite-element mesh must be refined in these regions if the discrete solution is to converge uniformly.

Classical adaptive finite-element codes use a posteriori residual estimators: solve on a coarse mesh, estimate the local error from the residual, refine, repeat. The approach works but requires a full solve per iteration. A complementary strategy is available once the metric is known analytically: use itself as an a priori refinement indicator, evaluated on vertices of the current mesh before any solve.

Definition (Curvature-CFL refinement criterion).

Fix a dimensionless tolerance . An edge of the mesh is marked for splitting if

and marked for collapsing if the same product is below a lower threshold with hysteresis constant .

The criterion [eq:curvature-cfl-criterion] is the geometric analogue of a CFL condition. It balances the local edge length against the local curvature scale, so that no edge spans more than units of "curvature-length". The cartan-remesh crate implements a criterion of the same form, with the larger principal curvature of a surface, in adaptive_remesh, which runs split and collapse passes on flat triangle meshes. For tetrahedral meshes the crate has indicator-driven barycentric refinement (refine_to_depth, non-conforming on faces shared with an unflagged neighbour) and uniform conforming red refinement. Conforming adaptive refinement of tetrahedra driven by remains to be written.

For anisotropic equation [eq:R-scalar] generalises. The scalar curvature of on is built from first and second derivatives of contracted with and is computable symbolically once and for all. For the transversely isotropic reservoir model , the expression is tractable in closed form. For fully anisotropic , it is evaluated numerically from and its first two derivatives using finite differences on the mesh. Either way the work per vertex is and needs no solve, whereas an a posteriori residual estimate needs a global solve first.

Remark (Length cross-ratio regularisation).

Scalar curvature determines where to refine. For triangle meshes cartan-remesh also provides length cross-ratio (LCR) conformal regularisation: a discrete spring energy on the cross-ratios of edge lengths in each pair of adjacent triangles, whose gradient flow moves vertices back towards a reference set of cross-ratios. LCR smoothing improves element shape without guaranteeing well-centred simplices. Positivity of a diagonal Hodge star must still be checked simplex by simplex, with the non-diagonal star of Remark [anisotropic-hodge] as the fallback. LCR has no tetrahedral counterpart in cartan-remesh.

7. Full pipeline

Assembling the pieces gives a seven-stage pipeline from microstructure to discrete solution.

Definition (RVE-to-mesh pipeline).

Given microstructural data for a heterogeneous continuum on a domain :

  1. Microstructure. Measure or model the phase volume fractions , the inclusion shape tensors and the phase conductivities .

  2. Homogenisation. Apply a scheme from section 3 (dilute, Mori-Tanaka or self-consistent) to produce as a section of the fibre bundle over .

  3. Metric. Set and . This gives a Riemannian manifold and the conformally related metric whose Laplace-Beltrami operator is Darcy's (Proposition [darcy-metric]).

  4. Initial mesh. Generate a coarse tetrahedral complex covering . cartan-dec's built-in generators (icosphere, torus, unit-square grid) produce surfaces and planar domains. A 3D reservoir mesh comes from an external tetrahedral mesher and enters through Mesh::from_simplices_generic.

  5. Discretisation. Assemble and in the metric (diagonal where usable, Galerkin otherwise) and the stiffness matrix on .

  6. Refinement. Evaluate at each vertex of , apply the curvature-CFL criterion [eq:curvature-cfl-criterion] and refine the marked elements. Interpolate across topology changes via geodesic interpolation inside the fibre.

  7. Solution. Solve the discrete pressure equation on the refined mesh, with the source integrated over each Euclidean dual cell. Post-process for streamlines, well tests and effective-property benchmarks.

Stage 2 is implemented in cartan-homog (dilute, Mori-Tanaka, self-consistent and Ponte Castañeda-Willis schemes). Stage 3 is a inverse and determinant. For stage 5 the Hodge stars of Remark [anisotropic-hodge] sit in cartan-mimetic and cartan-matfree. Stage 6 on tetrahedra is limited to what cartan-remesh provides today (section 6). Stage 7 needs a sparse symmetric solver such as preconditioned conjugate gradient.

8. Worked example of a transversely isotropic fractured reservoir

As a minimal end-to-end instance consider a fractured sandstone with matrix permeability and a single family of horizontal penny-shaped cracks with aspect ratio , density and orientation normal to . The crack density varies with depth on a characteristic scale .

Stage 1. The microstructural data is .

Stage 2. For penny-shaped cracks, the Mori-Tanaka effective permeability is transversely isotropic, , with

in the dilute limit (corrections of order from full Mori-Tanaka) for cracks far more conductive than the matrix (). The vertical correction is of order . The horizontal permeability oscillates with depth. The vertical permeability is unchanged because cracks normal to conduct horizontally only.

Stage 3. The metric is , varying with depth, and gives .

Stage 5. The Laplace-Beltrami operator of on vertex 0-forms is

with constant. It differs from Darcy's operator by the drift , as part (i) of Proposition [darcy-metric] predicts. The metric gives

which is Darcy's operator. The discrete form uses diagonal Hodge stars whose entries scale with : on horizontal edges and on vertical edges. The same weights are the transmissibilities of a two-point flux scheme on a -orthogonal grid.

Stage 6. The scalar curvature is . The second-derivative term dominates at small . Hence is largest where bends most sharply: at its crests and troughs, the zero-crossings of . The curvature-CFL criterion refines the mesh in horizontal slabs centred on those surfaces. The resulting mesh is dense at the crests and troughs of the crack density and sparse near its inflection surfaces.

From four numbers and one shape parameter the pipeline produces a discrete pressure solution on a mesh that reflects the depth structure of the fracture field. The refinement zones come from with no solve. Convergence of the discrete solution still has to be checked by refinement or by an a posteriori estimate.

9. Discussion

Closure of the pipeline. Each stage consumes the output of the previous one through a mathematical operation. Homogenisation produces a symmetric positive definite tensor, the algebraic type a Riemannian metric requires. The rescaled metric determines the Hodge star, the Hodge star determines the discrete Darcy operator, and the derivatives of determine the scalar curvature that drives mesh adaptation. No stage invents structure that the next stage ignores, and no stage asks for structure that the previous stage failed to provide.

Role of . Symmetric positive definite matrices are the right algebraic object for linear anisotropic response in any setting where energy is positive. The space is a Riemannian manifold in its own right, with the affine-invariant metric that respects congruence. When this structure is respected (geodesic interpolation in the fibre, affine-invariant distances for mesh comparison, Fréchet means for uncertainty quantification), every downstream operation inherits good properties: positive-definiteness preservation, invariance under coordinate rotation, stable interpolation under rank-one updates. Violating the SPD structure produces non-physical tensors under mesh refinement. Euclidean interpolation of entries with convex weights stays SPD but inflates the determinant (the swelling effect), and higher-order interpolation with negative weights can produce negative eigenvalues.

Omitted topics. Nonlinear homogenisation (where depends on the local field), time-dependent homogenisation (ageing viscoelasticity, where the Hill tensor acquires a memory kernel) and stochastic homogenisation (random microstructure, where is itself a random tensor with a limiting distribution) all sit outside the static linear pipeline described here. Each extension preserves the backbone of the present picture and adds one layer: a fixed-point iteration for nonlinear, a convolution in time for viscoelastic, a mean and covariance on for stochastic. The cartan workspace is structured to accommodate these as future crates without disturbing the flat linear case.

Connection to the wider programme. The RVE-to-mesh pipeline is one instance of a pattern that recurs in reservoir geometry, DEC numerical analysis and the active-matter and membrane work developed elsewhere on this site. In each case a heterogeneous physical system is described by a tensor field on a base manifold; the tensor field is interpreted as a geometric object (metric, connection, curvature); that geometry is discretised by DEC; and the discretisation is refined adaptively using an analytic curvature quantity.

References

  1. G. W. Milton. The Theory of Composites. Cambridge University Press, 2002.
  2. A. Zaoui. "Continuum micromechanics: survey." Journal of Engineering Mechanics, 128(8):808–816, 2002.
  3. J. D. Eshelby. "The determination of the elastic field of an ellipsoidal inclusion, and related problems." Proceedings of the Royal Society A, 241:376–396, 1957.
  4. P. Ponte Castañeda and J. R. Willis. "The effect of spatial distribution on the effective behavior of composite materials and cracked media." Journal of the Mechanics and Physics of Solids, 43(12):1919–1951, 1995.
  5. A. Bensoussan, J.-L. Lions, and G. Papanicolaou. Asymptotic Analysis for Periodic Structures. North-Holland, Amsterdam, 1978.
  6. G. Allaire. "Homogenization and two-scale convergence." SIAM Journal on Mathematical Analysis, 23(6):1482–1518, 1992.
  7. A. N. Hirani. Discrete Exterior Calculus. PhD thesis, California Institute of Technology, 2003.
  8. M. Desbrun, A. N. Hirani, M. Leok, and J. E. Marsden. "Discrete exterior calculus." arXiv:math/0508341, 2005.
  9. X. Pennec, P. Fillard, and N. Ayache. "A Riemannian framework for tensor computing." International Journal of Computer Vision, 66(1):41–66, 2006.
  10. J.-F. Barthélémy. ECHOES: Extended Calculator of HOmogEnization Schemes (v1.0.0). Zenodo, 2026. zenodo.org/records/19679057 (doi:10.5281/zenodo.19679057)
  11. G. Nguetseng. "A general convergence result for a functional related to the theory of homogenization." SIAM Journal on Mathematical Analysis, 20(3):608–623, 1989.