MeanFieldHomogenization.jl

Appearance

Adding a homogenization scheme ​

Which cell does your scheme serve?

homogenize accepts any AbstractHomogenizationCell — an RVE or a Laminate. Schemes stay typed on the concrete cell they serve (_evaluate(rve::RVE, ::MyScheme, …)), so a scheme applied to a cell it does not serve falls through to the error-raising fallback instead of dispatching somewhere wrong. Type the scheme argument too: leaving it untyped is ambiguous with that fallback.

MeanFieldHomogenization.Schemes ships Voigt/Reuss, dilute (direct and dual), Mori-Tanaka, Maxwell, Ponte-Castañeda–Willis, self-consistent (symmetric and asymmetric) and differential, each in an elastic and an ageing-viscoelastic flavor. A new scheme slots in beside them:

  1. Create src/Schemes/<scheme_name>.jl and include it from src/Schemes/Schemes.jl.

  2. Declare the scheme type (struct MyScheme <: HomogenizationScheme) in src/Schemes/scheme_types.jl, and register its symbol alias in SCHEME_ALIAS (src/Schemes/homogenize.jl) if you want homogenize(rve, :myscheme, :C) to work.

  3. Implement _evaluate(rve, ::MyScheme, ::Val{p}; kw...). Provide one method per tensor order — dispatch on the order of reference_property(rve, scheme, p) — so the scheme serves elasticity and transport from one implementation, as _mt_dispatch does in mori_tanaka.jl.

  4. Go through contribution_helpers.jl. Never call stiffness_contribution and friends directly on a phase geometry: the _phase_* helpers apply the amount, honor the symmetrize setting, hand the kernel the correctly pre-projected reference medium, and branch on is_homogeneous_inclusion. Two invariants stated in that file's header must hold in your kernel too — all helpers of a given evaluation must share the same projected P₀, and orientation averaging must never be folded into a tensor product.

  5. Prefer the bundled seams (_phase_dilute_and_contribution, _phase_dilute_and_stress_average, _phase_compliance_and_contribution) when you need two objects per phase: they share the single expensive solve.

  6. Consider what your kernel requires of an inclusion. Needing only the contribution tensors keeps the scheme usable with inclusions entered through gate C of the inclusion contract; needing the concentration tensor restricts it.

  7. Export through src/Schemes/Schemes.jl and re-export from src/MeanFieldHomogenization.jl.

  8. Add a unit test under test/Schemes/ and a section in docs/src/manual/schemes.md.

A scheme need not act on an RVE

Two cells other than RVE carry scheme kernels, and both follow the same split: the scheme type is declared in src/Schemes/scheme_types.jl (so the hierarchy and SCHEME_ALIAS stay in one place) while its _evaluate method lives with the cell it acts on — Laminated in src/Laminates/, and the two N-body schemes ClusterModel / EquivalentInclusion in src/Assemblies/. If your scheme needs microstructural information an RVE does not carry — positions, in the N-body case — add a cell rather than widening RVE, and remember that MFH Studio discovers every concrete HomogenizationScheme automatically: a scheme the interface cannot drive must be listed in ASSEMBLY_SCHEMES (tools/mfhstudio/mfhstudio/model.py) so it is filtered out of its catalog.