Date: 2026-01-16 Module: src/homology Purpose: Comprehensive usage guide for BELFEM's computational topology module

Revision History:

Date	Changes
2026-01-16	Initial documentation
2026-01-16	Applied external review feedback: Added Critical Constraints section (mesh mutation, MPI, memory ownership, thread safety, topological vs FEM correctness, mesh requirements, scratch-field usage); Added Quick-Start section; Added comprehensive Glossary; Moved Common Pitfalls to top with 10 concrete examples
2026-08-31	Marked the manifold-filtering description as a design of record with no implementation: manifold_filter_3d() and check_surface() exist nowhere in the tree. Added a design note under "Manifold Cleanup"; corrected the glossary, the Internal Workflow step, and the troubleshooting entry that told readers to tune thresholds that do not exist. Named the shipped path, Cohomology::clean_spfa() → remove_cut_pockets(). Content status remains Gregory Giard's to confirm

Module Contracts

Before diving into usage details, understand these fundamental design contracts:

Mesh Dependency: All topology operations require a valid, connected Mesh object
Integer Arithmetic: Homology computations use exact integer arithmetic (no floating-point errors)
Algorithm Independence: Results are topologically equivalent across algorithms; no performance comparison is recorded
Manifold Requirement: Thin cuts must form valid 2-manifolds for FEM assembly
MPI Awareness: Cohomology computation is currently serial (mesh must be on a single process)

Critical Constraints

⚠️ READ THIS SECTION BEFORE USING THE HOMOLOGY MODULE ⚠️

Mesh Mutation Contract

The homology module directly modifies mesh topology in invasive, irreversible ways:

Operations performed:

Duplicates nodes on cut surfaces (creates new Node objects with new IDs)
Relinks elements to point to duplicated nodes (modifies Element connectivity arrays)
Creates new SideSets for cut surfaces (adds to mesh's sideset collection)
Creates abstract nodes not belonging to any element (new DOF carriers)
Calls mesh->unfinalize() and mesh->finalize() multiple times during processing

Consequences:

NOT REVERSIBLE: Cannot undo node duplication or element relinking
Invalidates indices: Any code relying on original node indices will break
Invalidates fields: Custom mesh fields indexed by original nodes become inconsistent
Changes mesh size: Node count, element connectivity, and sideset count all increase

Execution order requirement:

// CORRECT order
Mesh* tMesh = new Mesh("problem.exo");
mesh::CutFactory tFactory(tMesh, ...);
tFactory.run();                          // Mutates mesh
fem::DofManager* tDofManager = new fem::DofManager(tMesh);  // AFTER cuts
 
// WRONG order
fem::DofManager* tDofManager = new fem::DofManager(tMesh);  // BEFORE cuts
tFactory.run();  // ❌ DOF manager now has invalid node references

Cannot run twice:

// WRONG
tFactory.run();  // First run succeeds
tFactory.run();  // ❌ UNDEFINED BEHAVIOR - mesh already mutated

MPI Serial Limitation

CRITICAL: Cohomology computation is serial-only and must run on an undistributed mesh.

What breaks if you ignore this:

Running on partitioned mesh → incorrect simplicial complexes (missing connectivity across ranks)
Running on distributed mesh → incomplete generators (topological holes span process boundaries)
Running on multiple ranks → rank-dependent results (nondeterministic cut generation)

Correct MPI workflow:

// 1. Load mesh on master rank ONLY
Mesh* tMesh = nullptr;
if (comm_rank() == 0) {
    tMesh = new Mesh("problem.h5");
 
    // 2. Run homology on master (serial)
    mesh::CutFactory tFactory(tMesh, ...);
    tFactory.run();
}
 
// 3. Partition AFTER homology
if (comm_rank() == 0) {
    tMesh->partition(comm_size());
}
 
// 4. Distribute to all ranks
if (comm_rank() == 0) {
    tMesh->distribute();
} else {
    tMesh = new Mesh();
    tMesh->receive(0);
}

Future work: Parallel cohomology is planned but not yet implemented.

Memory Ownership

Critical ownership rules:

Object	Created By	Owned By	Caller Action
SideSet* (cuts)	CutFactory	Mesh	Do NOT delete
Node* (abstract)	CutFactory	Mesh	Do NOT delete
Node* (duplicates)	CutFactory	Mesh	Do NOT delete
Cochain* (generators)	Cohomology	Cohomology	Do NOT delete
Chain* (generators)	Homology	Homology	Do NOT delete
Protoshell*	User	User	Caller must delete

Key rule: CutFactory transfers all created mesh entities (nodes, sidesets) to the mesh. The mesh takes ownership and will delete them in its destructor.

// CORRECT
CutFactory tFactory(tMesh, ...);
tFactory.run();
Cell<SideSet*>& cuts = tFactory.cuts();  // Reference only
// cuts are owned by tMesh - do NOT delete
 
delete tMesh;  // Mesh destructor deletes cuts and nodes
 
// WRONG
Cell<SideSet*>& cuts = tFactory.cuts();
delete cuts(0);  // ❌ DOUBLE DELETE when mesh destructor runs

Thread Safety

The homology module is NOT thread-safe:

SimplicialComplex::reduce_*() modifies internal state (chain/cochain maps)
CutFactory::run() modifies mesh topology (node duplication, element relinking)
smithForm() and friends modify their input matrix in place (aMat is taken by reference)

OpenMP: Do not call homology functions from parallel regions

MPI: Must run on single rank (see MPI Serial Limitation above)

Topological vs FEM Correctness

Critical distinction: The homology module provides two separate guarantees:

Topological Guarantees

✅ What homology guarantees:

Generators span correct cohomology groups H^k
Algorithms produce topologically equivalent results (different bases, same groups)
Betti numbers are exact (integer arithmetic, no rounding errors)
Orientation is consistent after cleanup

FEM Guarantees

⚠️ What homology does NOT automatically guarantee:

Thin cuts form valid 2-manifolds → Requires manifold filtering (may fail)
Node duplication yields independent DOFs → Requires proper element relinking (usually works)
No element has mixed φ⁺/φ⁻ nodes → Requires correct cut-side detection (usually works)
Static condensation produces correct constraints → Requires correct FEM implementation (user responsibility)

Implication: A topologically correct cohomology generator may still produce a non-manifold thin cut that breaks FEM assembly. Always inspect the cut sidesets (debug output, ParaView) — there is no is_manifold() query on SideSet.

Mesh Requirements

Preconditions for correctness:

✅ Required mesh properties:

Mesh must be connected (single component, or separate components with distinct domain types)
Mesh must be orientable (no Möbius strips)
Elements must form a manifold (no hanging faces, no non-manifold edges)
Facet orientations must be computable (consistent normal directions)

⚠️ Topology requirements:

Phi/non-phi domains must be correctly tagged (see Topology::run())
Thin-shell interfaces must be properly labeled as sidesets
Terminals (if used) must exist as nodes in the mesh

❌ Unsupported configurations:

Non-orientable manifolds (Möbius strip, Klein bottle)
Meshes with dangling tetrahedra (elements connected only at nodes/edges)
Multiply-connected phi-domains where cuts span >10^7 faces (memory limit)

Validation pattern:

// Check basic mesh properties
BELFEM_ERROR(tMesh->number_of_elements() > 0, "Empty mesh");
BELFEM_ERROR(tMesh->number_of_blocks() > 0, "No blocks defined");
 
// Check topology analysis succeeded
mesh::Topology tTopology(tMesh);
tTopology.run();
BELFEM_ERROR(tTopology.phi_block_ids().size() > 0,
             "No phi-domain blocks identified");

Scratch-Field Usage

Warning: the module uses the mesh entities' flag bits (flag_*() / unflag_*()) as scratch and, in the SPFA infeasibility path, writes Element::level (cl_Cohomology.cpp:510-528). It does not write owner or index.

Consequences:

Do not rely on entity flags or Element::level after CutFactory::run()
Run homology before partitioning because cohomology needs the undistributed mesh (see "MPI Serial Limitation"), not because of owner

// WRONG order
tMesh->partition();  // Distributes the mesh
tFactory.run();      // ❌ Cohomology needs the undistributed mesh
 
// CORRECT order
tFactory.run();      // Serial mesh
tMesh->partition();  // Now distribute

Quick-Start

Minimal example to generate cuts for a toroidal mesh:

#include "cl_CutFactory.hpp"
#include "cl_Topology.hpp"
 
// Load mesh
Mesh* tMesh = new Mesh("toroid.exo");
 
// Analyze topology
mesh::Topology tTopology(tMesh);
tTopology.run();
 
// Generate cuts
mesh::CutFactory tFactory(
    tMesh, &tTopology, {},  // Empty protoshells
    mesh::CutAlgorithm::PellikkaGeneralized
);
tFactory.run();
 
// Use results
for (mesh::SideSet* cut : tFactory.cuts()) {
    cut->set_domain_type(mesh::DomainType::Cut);
}

That's it! The tFactory.cuts() are now ready for FEM assembly with jump conditions.

Glossary

Core Concepts:

Term	Definition
Chain	Formal sum of k-simplices with integer coefficients (homology side): c = Σ aᵢ σᵢ where aᵢ ∈ ℤ
Cochain	Dual to chain; element of cochain complex (cohomology side): φ : C_k → ℤ
Thick Cut	Cohomology generator as a set of directed edges with ±1 coefficients (topological object)
Thin Cut	FEM-compatible surface where φ is discontinuous (minimal 2-manifold derived from thick cut). Worked example: thick_thin_cuts_and_conjugate_edges.md
Conjugate Edges	Face-loop edges bounding the thin cut, dual to the apex edges the thick cut intersects (see thick_thin_cuts_and_conjugate_edges.md)
Abstract Node	Node not belonging to any element; carries DOF for jump condition [φ] = φ⁺ - φ⁻
Duplicate Node	Copy of original node created during cut processing; allows φ⁺ ≠ φ⁻ on cut surface
Simplicial Complex	Collection of simplices (nodes, edges, faces, elements) with incidence relations
Boundary Operator	Map ∂_k : C_k → C_{k-1} (takes k-chain to its (k-1)-boundary)
Coboundary Operator	Map δ^k : C^k → C^{k+1} (dual to boundary operator)
Smith Normal Form	Diagonal matrix decomposition QAR = D revealing Betti numbers and torsion
Betti Number	Rank of k-th homology/cohomology group; counts k-dimensional "holes"
Homology Group	H_k = Ker(∂_k) / Im(∂_{k+1}) (k-cycles modulo k-boundaries)
Cohomology Group	H^k = Ker(δ^k) / Im(δ^{k-1}) (k-cocycles modulo k-coboundaries)
Terminal	Electrical connection point (set of nodes) for current injection/extraction
Protoshell	Configuration object for thin shell structures (sidesets, terminals, material, thickness)
Phi-domain	Region where magnetic scalar potential φ is solved (non-conducting)
Non-phi-domain	Conducting region or external boundary (not part of φ solution domain)
Manifold	Surface where every edge belongs to exactly 2 faces (topologically valid)
Non-manifold	Surface with edges shared by >2 faces or other defects (invalid for FEM)
Double Pocket	Topological defect: small cluster of faces attached via single articulation point

Algorithm Terms:

Term	Definition
Elementary Collapse	Removal of k-simplex and its free (k-1)-face (preserves homology)
Interior Face Reduction	Merging two k-simplices sharing a common face
Generalized Combine	Enhanced reduction updating boundaries after removing shared faces
Pellikka Algorithm	Original reduction via elementary collapses and combines
PellikkaGeneralized	Enhanced Giard algorithm combining CCR with Pellikka
CCR	Chain Complex Reduction (Kaczynski); classical acyclic matching
BeltedTree	Spanning tree approach with belt fasteners for cohomology
Manifold Filtering	Post-processing to remove non-manifold artifacts (double pockets, etc.). Design of record — not implemented; see the design note under "Manifold Cleanup"
Tarjan's Algorithm	DFS-based articulation point detection for manifold cleanup. Exists only as an archived prototype, archive/graph/fn_Graph_tarjan.{hpp,cpp}, which is outside the build

FEM Integration:

Term	Definition
Jump Condition	Discontinuity enforced on cut: [φ] = φ⁺ - φ⁻ = I (Ampere current)
Static Condensation	FEM technique to eliminate internal DOFs via transformation matrix T
Poincaré-Lefschetz Duality	Theorem allowing thick-to-thin cut conversion (edge set → surface)
h-φ Formulation	Magnetostatic FEM using magnetic field H = -∇φ in non-conducting regions
Ampere's Law	∮ H·dl = I for any loop encircling conductor with current I

Common Pitfalls

⚠️ Read this section to avoid frequent mistakes.

1. Building the CutFactory Before the Topology Has Run

Problem: Topology not analyzed before creating CutFactory.

// ❌ WRONG
mesh::Topology tTopology(aMesh);
mesh::CutFactory tFactory(aMesh, &tTopology, {});
tFactory.run();  // ❌ Topology::run() was never called!
 
// ✅ CORRECT
mesh::Topology tTopology(aMesh);
tTopology.run();  // Analyze FIRST
mesh::CutFactory tFactory(aMesh, &tTopology, {});
tFactory.run();

Consequence: phi_block_ids() is empty → no cohomology generators → no cuts.

2. Running Homology After Mesh Partitioning

Problem: Partitioning before homology breaks simplicial complex.

// ❌ WRONG
tMesh->partition(comm_size());  // Distributes across ranks
tFactory.run();  // ❌ Mesh is partitioned - cohomology sees incomplete topology
 
// ✅ CORRECT
tFactory.run();  // Run on full mesh FIRST
tMesh->partition(comm_size());  // Partition AFTER

Consequence: Incomplete generators, missing cuts, or crashes due to missing connectivity.

3. Deleting Cuts or Abstract Nodes

Problem: Manually deleting mesh entities owned by the mesh.

// ❌ WRONG
Cell<SideSet*>& cuts = tFactory.cuts();
delete cuts(0);  // ❌ DOUBLE DELETE - mesh owns this!
 
// ✅ CORRECT
Cell<SideSet*>& cuts = tFactory.cuts();
// Just use them - mesh will delete when tMesh destructor runs

Consequence: Double delete → crash or memory corruption.

4. Forgetting Edge/Face Creation

Problem: SimplicialComplex needs edges and faces, but mesh doesn't have them.

// ❌ WRONG
Mesh* tMesh = new Mesh("mesh.msh");
mesh::SimplicialComplex tComplex(tMesh);  // ❌ No edges!
 
// ✅ CORRECT
Mesh* tMesh = new Mesh("mesh.msh");
tMesh->create_edges();  // Create edges first
tMesh->create_faces();  // Create faces (for 3D)
mesh::SimplicialComplex tComplex(tMesh);

Consequence: number_of_ksimplices(1) == 0 → cohomology computation fails.

5. Mixing Chain and Cochain

Problem: Using Chain where Cochain expected (or vice versa).

// CutProcessor takes the whole Cohomology object in its constructor
// (cl_CutProcessor.hpp:94-101); it is never fed individual Chain or Cochain
// pointers. Homology generators are Chain*, cohomology generators are
// Cochain* — do not pass one where the other is expected, e.g.
// Cohomology::updatekGeneratorsFromHomology() takes Cell<Chain*>&.
Chain*   tChain   = tHomology.get_Generators()( 1 )(0);    // homology
Cochain* tCochain = tCohomology.get_Generators()( 1 )(0);  // cohomology

Consequence: Compile error or incorrect cut generation.

6. Not Checking Manifold Status

Problem: Non-manifold cuts break FEM assembly.

// ⚠️ RISKY
tFactory.run();
// Assume all cuts are manifold
 
// ✅ SAFER (pseudo-code: SideSet has no edges() accessor; walk the facets' edges yourself)
tFactory.run();
for (SideSet* cut : tFactory.cuts()) {
    // Check manifold status
    uint nonManifoldEdges = 0;
    for (Edge* edge : cut->edges()) {
        if (edge->number_of_facets() != 2) {
            ++nonManifoldEdges;
        }
    }
    BELFEM_ERROR(nonManifoldEdges == 0,
        "Cut %lu has %u non-manifold edges",
        (luint)cut->id(), nonManifoldEdges);
}

Consequence: FEM assembly fails due to edges shared by >2 faces.

7. Using Wrong Algorithm for Mesh Size

Problem: Using CCR (the theory/debugging path) for production meshes.

// ❌ Theory/debugging path
mesh::CutFactory tFactory(..., mesh::CutAlgorithm::CCR);
tFactory.run();
 
// ✅ Production default
mesh::CutFactory tFactory(..., mesh::CutAlgorithm::PellikkaGeneralized);
tFactory.run();

Consequence: No timing comparison between the algorithms is recorded (see README, "Cut Algorithms").

8. Running CutFactory Twice

Problem: Calling run() multiple times on same factory.

// ❌ WRONG
tFactory.run();  // First run - mutates mesh
tFactory.run();  // ❌ UNDEFINED BEHAVIOR - mesh already modified
 
// ✅ CORRECT
tFactory.run();  // Run exactly once
// Create new factory if you need different cuts

Consequence: Crash, incorrect cuts, or memory corruption.

9. Ignoring Orphaned Nodes

Problem: Orphaned nodes increase memory usage and may leak.

// ⚠️ CHECK
tFactory.run();
if (tFactory.orphaned_nodes().size() > 0) {
    message(InfoLevel::Default, "Warning: %lu orphaned nodes",
            tFactory.orphaned_nodes().size());
    // Optionally: delete orphaned nodes if not needed
}

Consequence: Memory leak or unexpected behavior if orphaned nodes are referenced elsewhere.

10. Incorrect Terminal-to-Shell Mapping

Problem: Terminal indices don't match protoshell indices.

// ❌ WRONG
Cell<Cell<id_t>> tTerminals = {{101}, {201}};
Cell<id_t> tShellIndices = {1, 0};  // Terminal 0 on shell 1? Mismatch!
 
// ✅ CORRECT
Cell<Cell<id_t>> tTerminals = {{101}, {201}};
Cell<id_t> tShellIndices = {0, 0};  // Both terminals on shell 0

Consequence: Terminals assigned to wrong thin shell → incorrect current distribution.

Introduction

The homology module provides computational topology tools for analyzing and processing finite element meshes, with a focus on electromagnetic field simulations in multiply-connected domains.

Primary use case: Generating topological cuts for enforcing current constraints in magnetostatic h-φ formulations.

What is a Cohomology Cut?

In multiply-connected domains (e.g., toroidal magnets), the magnetic scalar potential φ must satisfy Ampere's law:

∮_C H · dl = I

Where C is any closed loop encircling a conductor carrying current I. To enforce this in FEM:

Compute cohomology generator: Identifies topological "hole" in the mesh
Convert to thin cut: Minimal surface where φ is discontinuous
Duplicate nodes: Create separate DOFs for φ⁺ and φ⁻ on cut surface
Enforce jump condition: [φ] = φ⁺ - φ⁻ = I via static condensation

Location: src/homology/ CMake Target: belfem_homology

Module Architecture

File Organization

src/homology/
├── Entry Point
│   └── cl_CutFactory.{hpp,cpp}
│
├── Computational Core
│   ├── cl_Cohomology.{hpp,cpp}
│   ├── cl_Homology.{hpp,cpp}
│   ├── cl_SimplicialComplex.{hpp,cpp}
│   └── fn_Smith.{hpp,cpp}
│
├── Data Structures
│   ├── cl_Chain.{hpp,cpp}
│   ├── cl_Cochain.{hpp,cpp}
│   ├── cl_CutData.{hpp,cpp}
│   └── cl_CutSet.{hpp,cpp}
│
├── Processing
│   ├── cl_CutProcessor.{hpp,cpp}
│   ├── cl_CutProcessorManual.{hpp,cpp}
│   └── cl_InterfaceProcessor.{hpp,cpp}
│
├── Configuration
│   ├── cl_Topology.{hpp,cpp}
│   └── en_CutAlgorithm.hpp
│
└── Algorithms
    └── cl_BeltedTree.{hpp,cpp}

Dependencies

mesh: Mesh, Node, Element, Facet, SideSet classes
linalg: Matrix, Vector for boundary/coboundary matrices
containers: Cell, OrderedMap, Set for data structures
sparse: Integer matrix solvers for Smith Normal Form
comm: MPI support (computations currently serial)

CutFactory - Main Entry Point

Purpose

CutFactory is the primary user-facing class that orchestrates the entire cut generation pipeline.

File: cl_CutFactory.{hpp,cpp}

Constructor

CutFactory::CutFactory(
    Mesh*               aMesh,
    Topology*           aTopology,
    Cell<Protoshell*>&  aProtoshells,
    const CutAlgorithm  aAlgorithm,
    const bool          aUseEnrichment
);

Parameters:

aMesh: Finite element mesh to process
aTopology: Pre-analyzed mesh topology (identifies domains, interfaces)
aProtoshells: Configuration for thin shell structures
aAlgorithm: Cut generation algorithm (see Cut Algorithm Selection)
aUseEnrichment: Use XFEM enrichment (experimental)

Neither argument has a default; both are passed explicitly.

⚠️ Ownership: CutFactory does not take ownership of aMesh or aTopology. Caller must ensure they outlive the factory.

Key Methods

Workflow Methods

void set_terminals(
    const Cell<Cell<id_t>>& aTerminals,
    const Cell<id_t>&       aThinShellIndices
);

Sets up electrical terminal definitions for current injection/extraction points.

aTerminals: Cell of node ID lists, one per terminal
aThinShellIndices: Maps terminals to thin shell indices

void run();

Main execution pipeline. Performs:

Thin shell preprocessing
Cohomology computation
Thick-to-thin cut conversion
Node duplication
Manifold cleanup

⚠️ Call exactly once. Calling run() multiple times is undefined behavior.

Result Retrieval

Cell<Node*>& abstract_nodes();

Returns abstract nodes (new degrees of freedom) created for cut discontinuities.

Usage: Assign these to FEM DOF manager for φ⁺ and φ⁻ values.

Cell<SideSet*>& cuts();

Returns generated cut surfaces as SideSet objects.

Usage: Mark as DomainType::Cut and apply discontinuity constraints in FEM assembly.

Cell<Node*>& orphaned_nodes();

Returns nodes not belonging to any element after node duplication.

Usage: Typically freed after FEM setup, but may be used for debugging.

Vector<id_t>& thin_shell_boundaries();

Returns IDs of boundaries created for thin shell sidesets.

Cell<Facet*>& thin_shell_facets(const id_t aID);

Returns facets for a specific thin shell ID.

Debugging

void save_debug_meshes();

Saves one debug mesh per cut, named cut_<index>.vtk (cl_CutProcessor.cpp:132-140). It does not write separate thick-cut, thin-cut or manifold files — those filenames appear in older documentation but no code emits them.

Usage Pattern

#include "cl_CutFactory.hpp"
#include "cl_Topology.hpp"
 
// 1. Analyze mesh topology
mesh::Topology tTopology(aMesh);
tTopology.run();
 
// 2. Create CutFactory
mesh::CutFactory tFactory(
    aMesh,
    &tTopology,
    mProtoshells,
    mesh::CutAlgorithm::PellikkaGeneralized,
    false   // No enrichment
);
 
// 3. Set terminals (if needed)
Cell<Cell<id_t>> tTerminals = {{101, 102, 103}, {201, 202, 203}};
Cell<id_t> tShellIndices = {0, 0};
tFactory.set_terminals(tTerminals, tShellIndices);
 
// 4. Run pipeline
tFactory.run();
 
// 5. Extract results
for (mesh::SideSet* cut : tFactory.cuts()) {
    cut->set_domain_type(mesh::DomainType::Cut);
}
 
for (mesh::Node* node : tFactory.abstract_nodes()) {
    aDofManager->add_abstract_dof(node);
}

Cohomology Computation

Purpose

Cohomology computes cohomology groups H^k = Ker(δ^k) / Im(δ^{k+1}) from a simplicial complex.

File: cl_Cohomology.{hpp,cpp}

Constructor

// From simplicial complex
Cohomology::Cohomology(SimplicialComplex* aComplex, Mesh* aFullMesh);
 
// With belted tree
Cohomology::Cohomology(SimplicialComplex* aComplex, Mesh* aFullMesh, BeltedTree* aTree);

Parameters:

aComplex: Pre-constructed (and reduced) simplicial complex
aFullMesh: Original full mesh (for node/element lookups)
aTree: Belted tree for alternative algorithm (optional)

Both constructors leave the object with computed, cleaned generators, ready to use. The SimplicialComplex constructor computes all H^k groups; the BeltedTree constructor takes its H^1 generators from the tree.

Key Methods

void cohomologyGroupOfChainComplex();

Main computation method. Performs:

Kernel/image decomposition of the coboundary maps
Quotient group computation (Smith Normal Form)

Complex reduction happens beforehand on the SimplicialComplex; generator extraction is generatorsOfCohomology() and cleaning is clean(), both of which the constructor runs afterward.

⚠️ Called internally by the SimplicialComplex-based constructor. The BeltedTree constructor copies its generators from the tree instead. Do not call it yourself.

void quotientGroup();

Computes quotient space H^k = Ker(δ^k) / Im(δ^{k+1}) via Smith Normal Form.

Note: Typically called internally by cohomologyGroupOfChainComplex().

void generatorsOfCohomology();

Extracts cohomology generators from Smith decomposition.

Output: Populates mGenerators with Cochain objects representing basis elements.

void clean();

Cleans generator coefficients so every coefficient is in {-1, 0, 1}.

Method: Dispatches to clean_spfa(): SPFA feasibility certificate, greedy coboundary rectification, then remove_cut_pockets().

Note: Every constructor already calls clean(), so a constructed Cohomology always has unit-coefficient generators. CutFactory calls it a second time after updatekGeneratorsFromHomology(), because integer recombination of unit generators is generally non-unit again. Both runs are required: the second run is the only clean of the final generators, and the downstream CutData machinery can only represent coefficients in {-1, 0, 1}. See cohomology_algorithms.md, "Cleaning runs twice on the factory path".

Cell<Cell<Cochain*>>& get_Generators();

Returns cohomology generators organized by dimension k.

Structure: mGenerators[k] contains generators for H^k.

Cell<Matrix<int>>& get_CoboundaryMatrix();

Returns coboundary operator matrices δ^k.

Structure: mCoboundaryMatrix[k] is the δ^k : C^k → C^{k+1} matrix.

Visualization

void create_kGeneratorsField(uint k, Mesh* aMesh, string aLabel);

Creates mesh fields to visualize cohomology generators.

Parameters:

k: Dimension (0 for nodes, 1 for edges, 2 for faces)
aMesh: Mesh to attach fields to
aLabel: Field name prefix

Output: Creates fields showing generator support with ±1 coefficients.

Usage Pattern

#include "cl_Cohomology.hpp"
 
// Build the simplicial complex on the flagged mesh entities and reduce it
mesh::SimplicialComplex tComplex(aFullMesh, true);
tComplex.coreduce_complexPellikkaGeneralized();
 
// The constructor computes the cohomology groups and cleans the generators
mesh::Cohomology tCohomology(&tComplex, aFullMesh);
 
// Extract H^1 generators (most common for electromagnetic cuts)
Cell<Cochain*>& generators = tCohomology.get_Generators()( 1 );
 
message(InfoLevel::Default, "Found %d cohomology generators", generators.size());
 
// Visualize generators
tCohomology.create_kGeneratorsField(1, aFullMesh, "H1_gen");
 
// Access individual generator
Cochain* gen = generators(0);
for (auto& [edge_id, coeff] : gen->get_simplices_map()) {
    message(InfoLevel::Verbose, "Edge %lu: coefficient %d", edge_id, coeff);
}

Homology Computation

Purpose

Homology computes homology groups H_k = Ker(∂_k) / Im(∂_{k+1}), the dual counterpart to cohomology.

File: cl_Homology.{hpp,cpp}

Use cases:

Relative homology for terminal definitions
Orientation computation for current flow direction
Topological analysis of conductor geometries

Constructor

// With terminal suggestion
Homology::Homology(
    Mesh*               aMesh,
    Cell<Cell<id_t>>    aTerminals,
    Cell<id_t>          aThinShells
);
 
// From simplicial complex
Homology::Homology(SimplicialComplex* aComplex, Mesh* aFullMesh);

Key Methods

void homologyGroupOfChainComplex();

Main computation method. Performs:

Simplicial complex reduction
Quotient group computation (Ker/Im decomposition)
Generator extraction

void suggest_Homology(
    Cell<Cell<id_t>> aTerminals,
    Cell<id_t>       aThinShellIndices
);

Suggests homology generators based on terminal connectivity.

Use case: Pre-seeds homology computation with physically meaningful cycles.

Cell<int> generators_orientation(Cell<Vector<real>>& aVectors);

Computes orientation of generators relative to specified direction vectors.

Returns: Cell of ±1 values indicating generator orientation.

Use case: Ensures current flow direction matches physical conventions.

void reorient_generators();

Flips generator orientation based on computed directions.

Cell<Cell<Chain*>>& get_Generators();

Returns homology generators organized by dimension k.

Cell<Matrix<int>>& get_BoundaryMatrix();

Returns boundary operator matrices ∂_k.

Usage Pattern

#include "cl_Homology.hpp"
 
// Create with terminal hints
Cell<Cell<id_t>> tTerminals = {{101, 102}, {201, 202}};
Cell<id_t> tShellIndices = {0, 0};
 
mesh::Homology tHomology(aMesh, tTerminals, tShellIndices);
 
// Compute homology
tHomology.homologyGroupOfChainComplex();
 
// Get H_1 generators (1-cycles)
Cell<Chain*>& generators = tHomology.get_Generators()( 1 );
 
// Compute orientations
Cell<Vector<real>> tDirections = {{1, 0, 0}, {0, 1, 0}};
Cell<int> orientations = tHomology.generators_orientation(tDirections);
 
// Reorient if needed
tHomology.reorient_generators();

Cut Algorithm Selection

Available Algorithms

Four algorithms are available via the CutAlgorithm enum:

File: en_CutAlgorithm.hpp

enum class CutAlgorithm : uint
{
    Pellikka            = 0,  // Original Pellikka (2013)
    CCR                 = 1,  // Chain Complex Reduction (Kaczynski 2004)
    BeltedTree          = 2,  // Spanning tree approach
    PellikkaGeneralized = 3,  // Enhanced Giard - RECOMMENDED
    UNDEFINED           = 4
};

Algorithm Comparison

Algorithm	Complexity	Performance	Manifold Issues	Recommendation
PellikkaGeneralized	unmeasured	unmeasured	Rare	✅ Use this (production default)
Pellikka	unmeasured	unmeasured	Occasional	Legacy support
CCR	unmeasured	unmeasured	Frequent	Theory/debugging
BeltedTree	unmeasured	unmeasured	Rare	Alternative

Implementation Details

PellikkaGeneralized (Giard et al., in preparation):

Location: cl_SimplicialComplex.cpp — pGeneralizedCombine(), reduce_complexPellikkaGeneralized(), pGeneralizedCocombine(), coreduce_complexPellikkaGeneralized()
Method:
- Downward pass: pReduce(k) for k from d to 1 (elementary collapses)
- Second pass (again from d down to 1): interleaves pGeneralizedCombine(k) with pReduce(k-1); the cochain variant runs both passes upward, k = 0 … d−1
- For cohomology: Reversed dimensions (k from 0 to d-1)
Advantages:
- Near-linear complexity for large meshes
- Minimal Smith Normal Form computation
- Compact generator representations

Pellikka (Pellikka et al., 2013):

Location: cl_SimplicialComplex.cpp — pReduce(), pCombine(), reduceOmit()
Method:
- pReduce(): Elementary collapses (remove k-simplex and free (k-1)-face)
- pCombine(): Interior face reductions (merge neighboring simplices)
- reduceOmit(): Initialization pass
Advantages:
- Well-tested in electromagnetic community
- Widely used in GetDP, Sparselizard
- Stable for moderate mesh sizes

CCR (Kaczynski et al., 2004):

Location: cl_SimplicialComplex.cpp — reduce_complexCCR(), coreduce_complexCCR()
Method: Classical chain complex reduction via acyclic matching
Advantages:
- Mathematically pure approach
- Good for theoretical understanding
- Reliable for small meshes
Disadvantages:
- Slowest for large meshes
- May produce non-manifold cuts requiring extensive cleanup

BeltedTree:

Location: cl_BeltedTree.{hpp,cpp}
Method: Spanning tree construction with belt identification
Use case: Alternative approach when reduction methods struggle

Selection Guidance

// Production code - use generalized Pellikka
tFactory = new CutFactory(
    aMesh, aTopology, aProtoshells,
    CutAlgorithm::PellikkaGeneralized
);
 
// Legacy compatibility - original Pellikka
tFactory = new CutFactory(
    aMesh, aTopology, aProtoshells,
    CutAlgorithm::Pellikka
);
 
// Debugging/theory - CCR
tFactory = new CutFactory(
    aMesh, aTopology, aProtoshells,
    CutAlgorithm::CCR
);

Benchmark results: none recorded — see README, "Cut Algorithms".

Topology Analysis

Purpose

Topology analyzes mesh structure to identify domain types, interfaces, and boundary conditions.

File: cl_Topology.{hpp,cpp}

Constructor

Topology::Topology(Mesh* aMesh);

Key Methods

void run(); // and run_on_enriched_mesh() for a reloaded .bfm mesh

Performs mesh topology analysis:

Categorizes blocks by domain type (phi/non-phi)
Identifies interfaces between domains
Classifies sidesets by type
Builds connectivity maps

DomainType block_domain_type(const id_t aBlockID) const;

Returns domain type for a given block ID.

Domain types:

DomainType::Phi: Region where scalar potential φ is solved
DomainType::Conductor: Conducting region (source of magnetic field)
DomainType::Air: Air/vacuum region
DomainType::Boundary: Domain boundary
DomainType::Interface: Interface between domains
DomainType::Cut: Cohomology cut surface

const Cell<id_t>& phi_block_ids() const;

const Cell<id_t>& nonphi_block_ids() const;

Returns lists of block IDs categorized by domain type.

const Cell<id_t>& interface_ids() const;

Returns IDs of interface sidesets.

Usage Pattern

#include "cl_Topology.hpp"
 
mesh::Topology tTopology(aMesh);
tTopology.run();
 
// Check domain types
for (id_t blockID : tTopology.phi_block_ids()) {
    message(InfoLevel::Default, "Block %lu: Phi domain", blockID);
}
 
// Identify interfaces
for (id_t ssID : tTopology.interface_ids()) {
    SideSet* interface = aMesh->sideset(ssID);
    // Apply interface conditions
}
 
// Query specific block
if (tTopology.block_domain_type(5) == DomainType::Conductor) {
    // Apply source current density
}

Thin Shell Handling

Purpose

Protoshell configures thin shell structures for layered conductors (e.g., superconducting tapes, cables).

File: ../mesh/cl_Protoshell.hpp (header-only, namespace belfem)

Structure

class Protoshell
{
    Cell<id_t>   mSideSetIDs;       // Sidesets defining shell surfaces
    Cell<id_t>   mTerminalIDs;      // Terminal connection points
    Vector<real> mThicknesses;      // Shell thicknesses per layer
    Cell<string> mMaterials;        // Material names
    Cell<id_t>   mCurveIDs;         // Geometric curves
 
    // ... methods ...
};

Key Methods

void add_sideset(const id_t aID);
void add_terminal(const id_t aID);
void add_thickness(const real aValue);
void add_material(const string& aName);
void add_curve(const id_t aID);

Adds components to protoshell configuration.

const Cell<id_t>& sideset_ids() const;
const Cell<id_t>& terminal_ids() const;
const Vector<real>& thicknesses() const;

Accessors for configuration data.

Usage Pattern

#include "cl_Protoshell.hpp"
 
// Create protoshell configuration
mesh::Protoshell* tShell = new mesh::Protoshell();
 
// Add sidesets defining shell surfaces
tShell->add_sideset(101);  // Top surface
tShell->add_sideset(102);  // Bottom surface
 
// Add terminal connection points
tShell->add_terminal(201);
tShell->add_terminal(202);
 
// Add material and thickness
tShell->add_material("YBCO");
tShell->add_thickness(1e-6);  // 1 micron
 
// Pass to CutFactory
Cell<Protoshell*> tProtoshells = {tShell};
CutFactory tFactory(aMesh, &tTopology, tProtoshells, ...);

⚠️ Memory Management: CutFactory does not take ownership of Protoshell objects. Caller must delete them after use.

Cut Processing

Purpose

CutProcessor converts cohomology generators (thick cuts) to FEM-compatible thin cuts.

File: cl_CutProcessor.{hpp,cpp}

Key operations:

Converts thick cuts (edge sets) to thin cuts (face sets)
Duplicates nodes on cut surface
Relinks elements to create separate DOFs for φ⁺ and φ⁻
Creates abstract nodes for jump conditions
Performs manifold cleanup

Constructor

CutProcessor::CutProcessor(
    Mesh*               aMesh,
    Cohomology*         aCohomology,
    const Vector<id_t>& aPhiBlocks,
    const Vector<id_t>& aNonPhiBlocks,
    const Vector<id_t>& aPhiInterfaces,
    const Vector<id_t>& aShellBlocks
);

Parameters:

aMesh: Mesh to process
aCohomology: Cohomology computation results
aPhiBlocks: Block IDs for phi domains
aNonPhiBlocks: Block IDs for non-phi domains
aPhiInterfaces: Interface sideset IDs
aShellBlocks: Thin shell block IDs

Key Methods

Cell<Node*>& abstract_nodes();

Returns abstract nodes created for cut DOFs.

id_t max_node_id() const;

id_t max_sideset_id() const;

Returns maximum IDs after node duplication and cut creation.

Use case: Ensures unique IDs when creating additional mesh entities.

void save_debug_meshes();

Exports one debug mesh per cut for visualization, as cut_<index>.vtk (cl_CutProcessor.cpp:139-147).

Internal Workflow

The cut processing pipeline (called internally by CutFactory):

Thick cut extraction: Extract edges from cohomology generators
Face identification: Find faces incident to thick cut edges
Manifold filtering: Remove non-manifold artifacts (double pockets, etc.) — design of record; the step that actually runs is Cohomology::remove_cut_pockets(), see "Manifold Cleanup" below
Node duplication: Duplicate nodes on the positive side of the cut. Each cut set assigns copy ids while walking its originals in ascending id order, with the non-periodic originals first and the periodic pairs in the order encountered by that walk. The sets are processed in sorted order of their pattern hex strings. collect_duplicates then appends the abstract nodes and the copies in ascending copy-id order, so a copy's position never depends on the hash order of a duplicate map
Element relinking: Update element connectivity to duplicated nodes
Abstract node creation: Create DOFs for [φ] = φ⁺ - φ⁻
SideSet creation: Package cut as SideSet for FEM assembly

Manifold Cleanup

Design note — design of record, not wired in. manifold_filter_3d() and check_surface() do not exist anywhere in the tree, and no shipped code applies the three-phase scheme or the metric thresholds below. Tarjan articulation detection exists only as an archived prototype, archive/graph/fn_Graph_tarjan.{hpp,cpp}, which is outside the build — nothing in CMakeLists.txt references archive/.

What runs today is Cohomology::clean_spfa() → remove_cut_pockets() (cl_Cohomology.hpp:135,145), applied after SPFA rectification. Both are private — they run internally during cohomology computation and are not part of the caller-facing API. There is no CutProcessor method of either name.

The description is kept because it may be the design of record; it is marked so that no reader goes looking for the functions or tunes the thresholds. Its status is Gregory Giard's to confirm (doc/ai_collaboration_protocol.md §7.1). The fuller note is in cohomology_theory_and_implementation.md, "Post-Processing Cleanup".

Problem: Cohomology generators may produce non-manifold surfaces (edges shared by >2 faces).

Solution as designed: Multi-phase cleanup:

Phase 1: Tarjan's algorithm to detect articulation points (necks)
Phase 2: Region growing to detect embedded pockets
Phase 3: BFS layering with metric validation

Metrics as designed:

Cycle density > 0.3
Compactness < 2.0
Size < 20 faces

Implementation as shipped: Cohomology::clean_spfa() → remove_cut_pockets( const bool aFireTierA ) (cl_Cohomology.hpp:135,145). The designed filter above is not among the code paths that run.

Data Structures

Chain

Purpose: A chain is a formal sum of k-simplices with integer coefficients, keyed by the mesh entity's index (Node::index(), Edge::index(), Face::index() for the triangles of a 3D mesh, Element::index() for the top-dimensional cells), never by its id. It can also carry its own boundary and coboundary chains, which the simplicial-complex builder fills alongside it.

File: cl_Chain.{hpp,cpp}

Structure:

class Chain
{
    int                        mDim ;           // 0 nodes, 1 edges, 2 faces, 3 elements
    Mesh *                     mMesh ;
    OrderedMap< index_t, int > mSimplicesMap ;  // simplex index -> coefficient
    Chain *                    mBoundary ;      // (k-1)-chain, owned, or nullptr
    Chain *                    mCoboundary ;    // (k+1)-chain, owned, or nullptr
};

The constructor Chain( aDim, aMesh, aIsBound, aIsCobound ) allocates mBoundary unless aDim == 0 or aIsBound is set, and allocates mCoboundary unless aDim == 3 or aIsCobound is set. The two flags mark a chain that is someone's boundary or coboundary. A chain built with both flags false therefore owns a boundary (unless aDim == 0) and a coboundary (unless aDim == 3). Each of those owns at most one chain in the opposite direction (the boundary's coboundary and the coboundary's boundary), and those own nothing. The destructor deletes both members.

Key operations: operator+, operator-, and operator* modify this in place. operator() changes neither chain and returns the pairing. None of the operators returns a new chain.

void operator+( Chain * aChain ) ;        // this += aChain
void operator-( Chain * aChain ) ;        // this -= aChain
void operator*( const int aMult ) ;       // every coefficient *= aMult
int  operator()( Chain * aChain ) ;       // sum over shared indices of coeff * coeff
 
void add_simplex_to_chain( const index_t aIndex, const int aCoeff ) ;  // adds, erases a zero
void remove_simplex_from_chain( const index_t aIndex ) ;
void add_chain_to_chain( Chain * aChain, int aCoeff ) ;                // this += aCoeff * aChain
void remove_chain_from_chain( Chain * aChain ) ;
 
OrderedMap< index_t, int > & get_simplices_map() ;
int  get_coefficient( const index_t aIndex ) ;    // 0 when the index is absent
void set_coefficient( const index_t aIndex, const int aCoeff ) ;  // 0 erases the entry
int  get_dim() ;
 
Chain * get_boundary() ;
Chain * get_coboundary() ;
void add_simplex_to_boundary( index_t aIndex, const int aCoeff ) ;    // forwards to mBoundary
void add_simplex_to_coboundary( index_t aIndex, const int aCoeff ) ;  // forwards to mCoboundary
void print() ;

add_chain_to_chain refuses a dimension mismatch (it prints and returns). Otherwise, it also propagates into mBoundary and mCoboundary wherever this chain owns them, so a generator assembled from other chains keeps its boundary consistent. It adds the stored boundary and coboundary of aChain; it does not recompute them from the coefficients. Therefore, aChain must own every member that this chain owns: a chain built with a flag set has a null member in that direction. remove_chain_from_chain does not propagate.

Usage:

Chain tC1( 1, tMesh, false, false ) ;
tC1.add_simplex_to_chain( tEdgeA->index(),  1 ) ;
tC1.add_simplex_to_chain( tEdgeB->index(), -1 ) ;
 
Chain tC2( 1, tMesh, false, false ) ;
tC2.add_simplex_to_chain( tEdgeA->index(), 2 ) ;
 
tC1 + &tC2 ;                       // tC1 is now { A: 3, B: -1 }
int tPairing = tC1( &tC2 ) ;       // 3 * 2 = 6
 
for ( const auto & [ tIndex, tCoeff ] : tC1.get_simplices_map() )
{
    // tIndex is Edge::index(); look the edge up through the mesh, not by id
}

Cochain

Purpose: A Cochain is dual to Chain: a formal sum of k-cosimplices keyed by the same simplex indices. The generators that Cohomology hands out are Cochain *.

File: cl_Cochain.{hpp,cpp}

Structure: The structure is identical to Chain, but uses Cochain * boundary and coboundary members. The constructor and destructor follow the same rules.

Key operations: The in-place API corresponds to Chain, with cochain in the method names, with one difference: add_simplex_to_cochain and add_cochain_to_cochain return immediately for a zero coefficient. A zero-coefficient add_cochain_to_cochain therefore neither checks the dimension nor propagates. The Chain versions do not return early.

void operator+( Cochain * aCochain ) ;
void operator-( Cochain * aCochain ) ;
int  operator()( Chain * aChain ) ;       // evaluation <phi, c>: sum over shared indices
 
void add_simplex_to_cochain( const index_t aIndex, const int aCoeff ) ;
void remove_simplex_from_cochain( const index_t aIndex ) ;
void add_cochain_to_cochain( Cochain * aCochain, const int aCoeff ) ;
void remove_cochain_from_cochain( Cochain * aCochain ) ;
 
OrderedMap< index_t, int > & get_simplices_map() ;
int  get_coefficient( const index_t aIndex ) ;
void set_coefficient( const index_t aIndex, const int aCoeff ) ;
int  get_dim() ;
 
Cochain * get_coboundary() ;
Cochain * get_boundary() ;
void add_simplex_to_coboundary( index_t aIndex, const int aCoeff ) ;
void add_simplex_to_boundary( index_t aIndex, const int aCoeff ) ;
bool isBound() ;      // no callers in the tree
bool isCobound() ;
void print() ;

There is no scalar operator* on Cochain. add_cochain_to_cochain accumulates (this += aCoeff * aCochain), so to scale a cochain, add it with the factor to a new, empty cochain of the same dimension. Cohomology recombines generators this way into fresh cochains.

Usage:

Cochain tPhi( 1, tMesh, false, false ) ;
tPhi.add_simplex_to_cochain( tEdgeA->index(),  1 ) ;
tPhi.add_simplex_to_cochain( tEdgeB->index(), -1 ) ;
 
Chain * tC = ... ;
int tValue = tPhi( tC ) ;          // cochain evaluated on the chain

CutData

Purpose: Stores topology and metadata for a single cut.

File: cl_CutData.{hpp,cpp}

Structure:

class CutData
{
    Cell<id_t>           mEdgeIDs;        // Edges in thick cut
    Cell<id_t>           mElementIDs;     // Elements adjacent to cut
    Cell<Facet*>         mFacets;         // Facets forming thin cut
    DynamicBitset        mEdgeBitset;     // Fast edge membership test
    DynamicBitset        mNodeBitset;     // Fast node membership test
    DomainType           mType;           // Cut classification
};

Key methods:

void determine_cut_case(Element* aElement);

Determines how cut intersects an element (which faces are on cut).

Use case: Guides node duplication and element relinking.

bool node_is_on_cut(const id_t aNodeID) const;

bool edge_is_on_cut(const id_t aEdgeID) const;

Fast membership tests using bitsets.

SimplicialComplex

Purpose: Builds and reduces simplicial chain/cochain complexes.

File: cl_SimplicialComplex.{hpp,cpp}

Structure:

class SimplicialComplex
{
    Cell< Map< index_t, Chain * > >   mChainsMap;     // k-chains
    Cell< Map< index_t, Cochain * > > mCochainsMap;   // k-cochains
    Mesh *                            mMesh;          // the mesh the complex is built from
};

Key methods:

void create_complex(Mesh* aMesh, bool aPeriodicity);

Builds chain/cochain complex from flagged mesh entities.

Flags:

Nodes: 0-simplices
Edges: 1-simplices
Faces: 2-simplices
Elements: 3-simplices

void reduce_complexPellikkaGeneralized();
void reduce_complexPellikka();
void reduce_complexCCR();

Reduction algorithms (see Cut Algorithm Selection).

Cell<Matrix<int>> createMatrixFromBoundaryMap();

Cell<Matrix<int>> createMatrixFromCoboundaryMap();

Exports boundary/coboundary operators as integer matrices.

Structure: ∂_k : C_k → C_{k-1}, δ^k : C^k → C^{k+1}

Cut Sideset Creation

Purpose: Cohomology generators become mesh sidesets (the thin cuts).

Live path: CutFactory::run() calls compute_thin_cuts_and_duplicate_interface_nodes(), which constructs a CutProcessor. The CutProcessor constructor calls create_thin_cut_sidesets(), which delegates to CutData::add_thin_cut_sidesets_to_mesh(). Users do not call any of this directly; it runs inside CutFactory::run().

Note: The earlier SideSetFactory class and the CutFactory::create_sidesets_2d/3d methods that built sidesets directly from generator support were retired in 2026. CutProcessor/CutData is now the only sideset path.

InterfaceProcessor

Purpose: Handles node duplication and element relinking at domain interfaces (e.g., air-conductor boundaries).

File: cl_InterfaceProcessor.{hpp,cpp}

Structure:

class InterfaceProcessor
{
    Mesh* mMesh;
    Topology* mTopology;
    Cell<Node*>& mAbstractNodes;          // Output: abstract nodes
    Map<id_t, InterfaceSet*> mSetsMap;    // Interface sets by sideset ID
    DynamicBitset* mIsAirBlock;           // Air block flags
    DynamicBitset* mIsFerroBlock;         // Ferromagnetic block flags
};

Key methods:

InterfaceProcessor(
    Mesh* aMesh,
    Topology* aTopology,
    Cell<Node*>& aAbstractNodes,
    const uint aNumOriginalSideSets,
    const id_t aMaxNodeID
);

Constructor performs full interface processing pipeline:

Creates interface sets (groups of related interfaces)
Connects sidesets to sets
Duplicates nodes at interfaces
Relinks elements to duplicated nodes. A node is copied once per interface set, except on the conductor side. Each tie-weight set whose master block is a conductor takes its copy from one shared map keyed by the node's own id. A final sweep installs that copy in every element of every conductor block that still holds the original. Two conductor blocks that meet the air along a common line therefore reference the same copy, and the conductor edges leaving that line along their common face are created once. Without this sharing, each block had its own copy, so the edges were doubled and untied, and tangential h could jump across the air-adjacent row of the conductor|conductor face: a lossless sheet. Ferro, buffer and air-master sets keep private per-set copies; a ferro|ferro sideset stays split, as its material jump requires. On a periodic mesh, the copies of a periodic node pair are also tied to each other, either within one interface set or across two; this pairing runs at the end of node duplication, before the relink. Across sets, a pair is tied only when both sets have the same treatment and, for tie-weight sets, the same master domain type (pair_within). When no compatible set holds the partner's copy, because it lies only on a buffer|air, ferro|air, or air-master set against a conductor|air set, or only on a decoupled coil set against a tie-weight set, setup refuses the pair with a message naming both nodes and both domain types or treatments. The reason is the tie itself: it identifies like unknowns, h on conductor edges and φ elsewhere (Messe et al. 2023, §1), and cannot represent a change of formulation across the plane. The rule is applied to the set's master domain type, so it also refuses ferro against buffer, although both are φ blocks. A periodic plane that coincides with a change of domain type, or that a coil touches on one side only, can be avoided by shifting the plane to a section where the interface has one domain type and the same treatment on both sides. The plain "not duplicated in any interface set" refusal remains for a partner that lies on no interface set at all. Two rules come first. First, every interface sideset, that is, every sideset with a facet this processor duplicates at (one side air or ferro, not both air), must have one master domain type on all of its facets that have a slave; a sideset without such a facet is left alone. A sideset is admitted to an interface set, and the set is classed as shared or private and paired across a periodic plane, by its first facet, as in 0.9.2: a native Gmsh mesh builds one sideset per geometric entity, so its facets cannot differ in kind. refuse_mixed_master_types checks, after CutProcessor has run, every sideset that existed before it (generated cut sidesets are not examined, and thin-shell sidesets have no slave by then), before any sideset is admitted and before any interface copy exists. If it finds a mismatch, it reports the sideset id, the two domain types (name and enumerator value, since several types print under one name), the two facets with their master elements, and the remedy: split the sideset so that each piece has one master domain type. A conductor|air facet beside an air|air facet is refused the same way, since the air|air facet's master is air. A sideset over several blocks of one domain type stays accepted. A reused .bfm (one whose checksum and configuration tag match the deck) skips all of this: it replays the enrichment it was written with. Second, interface-copy numbering is a function of the mesh alone. Copies are allocated one set at a time, in set order (the sorted hex strings of the block-index bitsets of each set's block pair, a fixed function of the mesh's block order), and within each set in ascending order of the original id. A shared conductor copy is allocated by the first set that reaches the original. Copies are appended to the mesh in ascending copy-id order, which is the allocation order. Periodic pairs are backed up set by set: first, all pairs tied within one set, in ascending order of the original id; then the next set; finally, pairs tied across sets, again set by set and in ascending order. The maps that hold originals and copies are hash maps, but their iteration order never reaches the mesh. Therefore, interface-copy ids and positions are the same on every platform for the same incoming node order. The cut copies, which precede the interface copies in the mesh, are appended the same way: their ids come from a sorted walk of each cut set's originals (see "CutSet"), and CutProcessor::collect_duplicates sorts the kept copies by id after the abstract nodes before appending them. Therefore, no hash-map iteration order reaches a copy's position. Cut copies and interface copies occupy positions determined by the node order entering the cut processor. That is not yet a platform-independent node numbering. CutFactory renumbers the mesh with reverse Cuthill-McKee before the cohomology, and MaxwellFactory renumbers it again after enrichment, in the order that reaches the .bfm and the dof manager. symrcm sorts a node's neighbors by degree with std::sort, so equal degrees follow whatever order the C++ standard library's sort happens to leave them in. That behavior is reproducible on one toolchain but is not guaranteed across toolchains until the graph sort has a tie-break. The thin-shell copies were already appended in ascending index order of their originals. This does not cover a reused .bfm, which replays the numbering with which it was written, the dof manager's own ordering, or partitioned runs.
Ties each duplicate to its original (InterfaceTreatment::TieWeight1) or decouples it when a coil touches the interface (InterfaceTreatment::Decouple)

Helper class: InterfaceSet

class InterfaceSet
{
    Mesh* mMesh;
    DynamicBitset* mBitset;               // Interface pattern
    Cell<Element*> mElements;              // Elements on interface
    Map<id_t, Node*> mOriginals;          // Original nodes
    Map<id_t, Node*> mDuplicates;         // Duplicated nodes
    Cell<SideSet*> mSideSets;             // Sidesets in this set
};

Physical context: Duplicates the nodes of every sideset that separates a φ block (Air/Buffer/Ferro) from a non-φ block or Ferro (conductor-air, conductor-ferro, ferro-air), so the φ side and the other side carry distinct DOFs; coil-touching interfaces are decoupled (InterfaceTreatment::Decouple). It does not create abstract nodes; it receives the factory's list only to keep them out of the duplicate scan.

Usage:

// Internal use by CutFactory
Cell<Node*> tAbstractNodes;
mesh::InterfaceProcessor tProcessor(
    aMesh, &tTopology, tAbstractNodes,
    originalSidesetCount, maxNodeID
);
// tProcessor constructor performs all processing
// tAbstractNodes is only read (excluded from the duplicate scan)

Note: Runs unconditionally from CutFactory::compute_thin_cuts_and_duplicate_interface_nodes(), thin shells or not.

CutSet

Purpose: Manages node duplication for a single cohomology cut.

File: cl_CutSet.{hpp,cpp}

Structure:

class CutSet
{
    Mesh* mMesh;
    Cell<Node*>& mNodeOriginals;          // Candidate nodes for duplication
    DynamicBitset* mBitset;               // cut pattern: bit c set ⇔ this set's nodes are split by cut c
    DynamicBitset* mNodeBitset;           // Node switches (which side of cut)
    Map<id_t, Node*> mNodeDuplicates;     // Original → duplicate mapping
};

Key methods:

CutSet(
    Mesh* aMesh,
    Cell<Node*>& aNodeOriginals,
    const string& aHexString,          // Bitset pattern as hex
    const index_t aNumberOfCuts
);

Constructor parses hex-encoded bitset describing cut topology.

bool test(const Node* aNode) const;

Tests if node is on the cut surface.

Returns: true if node index is set in cut bitset.

void create_duplicates(id_t& aMaxNodeID, Cell<Node*>& aAbstractNodes);

Creates duplicate nodes for positive side of cut.

Algorithm:

Iterate through nodes on the cut in ascending id (CutProcessor::collect_nodes sorts the originals): the non-periodic originals first, then the periodic pairs in the order encountered by the walk
Create a new Node with a unique ID in that order
Store it in the mNodeDuplicates map (a hash map; its iteration order never reaches the mesh, because CutProcessor::collect_duplicates sorts the copies by id before appending them)
Tie each duplicate to its original plus the abstract nodes of the cuts in this set's pattern (weight 1 each; the abstract nodes are created by CutProcessor::create_abstract_nodes())

Node* duplicate(Node* aNode);

Retrieves previously created duplicate for a node.

Returns: Duplicate node pointer (must exist in map).

DynamicBitset* node_bitset();

Map<id_t, Node*>& duplicate_map();

Accessors for internal data structures.

Usage:

// Internal use by CutProcessor
mesh::CutSet tCutSet(
    aMesh, nodeOriginals, hexBitset, numCuts
);
 
// Test membership
if (tCutSet.test(someNode)) {
    // Node is on cut
}
 
// Create duplicates
tCutSet.create_duplicates(maxNodeID, abstractNodes);
 
// Access duplicate
mesh::Node* dup = tCutSet.duplicate(originalNode);

Bitset encoding: Uses DynamicBitset with hex string serialization for efficient cut pattern storage and transmission.

Note: This is an internal class used by CutProcessor during thick-to-thin conversion.

Algorithm Classes and Functions

BeltedTree

Purpose: Alternative algorithm for cohomology computation using spanning tree construction with belt identification.

File: cl_BeltedTree.{hpp,cpp}

Structure:

class BeltedTree
{
    Cell<Chain*> m1HomologyGenerators;       // H_1 generators (input)
    Cell<Cochain*> m1CohomologyGenerators;   // H^1 generators (output)
    Cell<Edge*> mBeltFasteners;              // Edges connecting tree to belts
    Cell<index_t> mTree;                     // Spanning tree edge indices
    SimplicialComplex* mSimplicialComplex;   // Reduced complex
    Mesh* mMesh;                             // Full mesh
};

Algorithm concept:

The belted tree method constructs cohomology generators by:

Building a spanning tree of the simplicial complex
Identifying "belts" (cycles not in the tree) corresponding to homology generators
Selecting "belt fasteners" (edges connecting tree to belts)
Computing dual cochains from the fastener configuration

Key methods:

BeltedTree(
    Mesh* aMesh,
    SimplicialComplex* aSimplicialComplex,
    Cell<Chain*> a1HomologyGenerators
);

Constructor requires pre-computed H_1 homology generators.

Precondition: Homology must be computed first.

void select_belt_fasteners();

Identifies edges that connect the spanning tree to independent cycles.

Algorithm: For each homology generator (1-cycle), find minimal edge set that completes the cycle when added to the tree.

void create_tree();

Constructs spanning tree of the 1-skeleton (edge graph).

Method: Breadth-first search or depth-first search from arbitrary root node.

void compute_cohomology();

Computes H^1 cochains from belt fastener configuration.

Duality: Each belt fastener corresponds to a cohomology generator via Poincaré duality.

Cell<Cochain*>& get_cohomology();

Returns computed cohomology generators.

Visualization:

void create_tree_field( Mesh * aEdgeMesh, const string aFieldName );

void create_cohomology_field( Mesh * aEdgeMesh );

Creates mesh fields to visualize spanning tree and cohomology generators.

Usage pattern:

// tComplex is a SimplicialComplex* built on the flagged mesh
// 1. Compute homology first (the constructor computes the groups)
mesh::Homology tHomology(tComplex, aFullMesh);
Cell<Chain*>& h1_gens = tHomology.get_Generators()( 1 );
 
// 2. Create BeltedTree
mesh::BeltedTree tTree(aMesh, tComplex, h1_gens);
 
// 3. Execute algorithm
tTree.select_belt_fasteners();
tTree.create_tree();
tTree.compute_cohomology();
 
// 4. Retrieve cohomology
Cell<Cochain*>& cohom_gens = tTree.get_cohomology();

When to use:

Alternative to reduction-based methods (Pellikka, CCR)
Useful when homology generators have specific physical meaning
Educational: Explicit geometric interpretation of duality

Comparison to other algorithms:

Advantage: Direct geometric construction (easy to visualize)
Disadvantage: Requires pre-computed homology (extra step)
Performance: unmeasured (no timing comparison recorded)

Selection in CutFactory:

mesh::CutFactory tFactory(
    aMesh, &tTopology, tProtoshells,
    mesh::CutAlgorithm::BeltedTree  // Use belted tree algorithm
);

fn_Smith - Smith Normal Form Functions

Purpose: Integer matrix decomposition for computing homology/cohomology groups.

File: fn_Smith.{hpp,cpp}

Mathematical background:

The Smith Normal Form decomposes an integer matrix A into:

Q * A * R = D

Where:

Q, R are unimodular matrices (det = ±1)
D is diagonal matrix with d_1 | d_2 | ... | d_r (divisibility chain)
Diagonal entries d_i are invariant factors
Number of nonzero entries r is the rank

Homology application:

Given boundary operator ∂_k : C_k → C_{k-1} with matrix representation A:

Ker(∂_k) dimension = n - r (n = number of k-chains)
Im(∂_{k+1}) dimension = r
H_k rank (Betti number) = dim(Ker) - dim(Im)
Torsion subgroup = {coefficients d_i where d_i > 1}

Core functions:

std::tuple<Matrix<int>, Matrix<int>, uint>

rowEchelon(Matrix<int>& aMat);

Computes row echelon form of integer matrix.

Returns:

Q: Row transformation matrix
Q_: Inverse of Q
rank: Number of nonzero rows

Algorithm: Euclidean-style elimination in exact integer arithmetic: pivot on the smallest non-zero entry, subtract integer quotients of rows (floor(a/b)), swap; no fractions ever appear.

std::tuple<Matrix<int>, Matrix<int>>

kernelImage(Matrix<int>& aMat);

Computes kernel and image of integer matrix.

Returns:

Kernel basis (columns span Ker(A))
Image basis (columns span Im(A))

Method: Row echelon form followed by basis extraction.

std::tuple<Matrix<int>, Matrix<int>, Matrix<int>, Matrix<int>, uint, uint>

smithForm(Matrix<int>& aMat);

Main Smith Normal Form decomposition.

Returns:

Q: Left transformation matrix
Q_: Inverse of Q
R: Right transformation matrix
R_: Inverse of R
rank: Number of nonzero diagonal entries
Number of rows/columns

Algorithm:

Row reduction to echelon form
Column reduction to eliminate above-diagonal entries
Iterative divisibility checking (ensure d_i | d_{i+1})
Minimize diagonal entries via elementary operations

Helper functions:

std::pair<uint, uint> minNonzero(Matrix<int>& aMat, const uint k);

Finds smallest nonzero entry in submatrix (optimization for faster reduction).

void moveMinNonzero(Matrix<int>& aMat, Matrix<int>& aQ, Matrix<int>& aQ_,

Matrix<int>& aR, Matrix<int>& aR_, const uint k);

Swaps rows/columns to move minimal element to pivot position.

std::tuple<bool, uint, uint, int>

checkForDivisibility(Matrix<int>& aMat, const uint k);

Checks whether the pivot B[k,k] divides every entry of the trailing submatrix B[k+1:end, k+1:end] (Kaczynski et al., Smith-form step); returns the first offending (row, col, quotient).

Returns: (divisible, row_index, col_index, violating_entry)

void partSmithForm(Matrix<int>& aMat, Matrix<int>& aQ, Matrix<int>& aQ_,

Matrix<int>& aR, Matrix<int>& aR_, const uint k);

Performs partial Smith form reduction for diagonal entry k.

Method: Euclidean algorithm to eliminate non-divisible entries.

Matrix<int> SolveInt(Matrix<int> aMat, Matrix<int>& aVec);

Solves integer linear system A*x = b using Smith decomposition.

Returns: Integer solution vector (if exists).

Usage in homology:

// Compute boundary matrix
Cell<Matrix<int>> boundary = tComplex->createMatrixFromBoundaryMap();
 
// Smith decomposition of ∂_k
auto [Q, Q_, R, R_, rank, size] = smithForm(boundary[k]);
 
// Extract homology generators
// Ker(∂_k) = columns of R corresponding to zero diagonal entries
// Betti_k = number of zero diagonal entries - rank(∂_{k+1})

Performance notes:

Complexity: O(n³) for dense matrices (n = dimension)
Optimization: Uses smallest-entry pivoting for faster convergence
Integer overflow: No protection for very large matrices (TODO: arbitrary-precision)
Sparsity: Does not exploit sparsity (future improvement)

Elementary operations:

The following operations preserve equivalence and update transformation matrices:

rowExchange(i, j): Swap rows i and j
columnExchange(i, j): Swap columns i and j
rowMultiply(i): Multiply row i by -1
columnMultiply(i): Multiply column i by -1
rowAdd(i, j, q): Add q times row j to row i
columnAdd(i, j, q): Add q times column i to column j

Corresponding operations (with transformation tracking):

rowExchangeOperation(): Updates Q, Q_, and matrix
columnExchangeOperation(): Updates R, R_, and matrix
rowAddOperation(): Updates Q, Q_, and matrix
columnAddOperation(): Updates R, R_, and matrix

Example:

#include "fn_Smith.hpp"
 
// Boundary matrix for 2-dimensional complex
Matrix<int> boundary2(4, 6);
// ... populate with boundary coefficients ...
 
// Compute Smith Normal Form
auto [Q, Q_, R, R_, rank, dim] = smithForm(boundary2);
 
// Betti number: number of zero diagonal entries
uint betti = dim - rank;
 
message(InfoLevel::Default, "H_2 Betti number: %u", betti);
 
// Extract generators
Matrix<int> generators(R.n_rows(), betti);
uint col = 0;
for (uint i = rank; i < dim; ++i) {
    for (uint row = 0; row < R.n_rows(); ++row) {
        generators(row, col) = R(row, i);
    }
    col++;
}

Thread safety: ⚠️ Not thread-safe on shared inputs: modifies the input matrix in place.

Common Patterns

Pattern 1: Basic Cut Generation

#include "cl_CutFactory.hpp"
#include "cl_Topology.hpp"
 
void generate_cuts(Mesh* aMesh)
{
    // 1. Analyze topology
    mesh::Topology tTopology(aMesh);
    tTopology.run();
 
    // 2. Create empty protoshell list (no thin shells)
    Cell<mesh::Protoshell*> tProtoshells;
 
    // 3. Create factory
    mesh::CutFactory tFactory(
        aMesh,
        &tTopology,
        tProtoshells,
        mesh::CutAlgorithm::PellikkaGeneralized
    );
 
    // 4. Run pipeline
    tFactory.run();
 
    // 5. Process results
    for (mesh::SideSet* cut : tFactory.cuts()) {
        cut->set_domain_type(mesh::DomainType::Cut);
        message(InfoLevel::Default, "Cut %lu: %u facets",
                cut->id(), cut->number_of_facets());
    }
}

Pattern 2: Terminal-Driven Cuts

void generate_terminal_cuts(Mesh* aMesh)
{
    // Topology analysis
    mesh::Topology tTopology(aMesh);
    tTopology.run();
 
    // Define terminals (node IDs)
    Cell<Cell<id_t>> tTerminals;
    tTerminals(0) = {101, 102, 103};  // Terminal 1
    tTerminals(1) = {201, 202, 203};  // Terminal 2
 
    // Map to thin shell indices (0 = not on thin shell)
    Cell<id_t> tShellIndices = {0, 0};
 
    // Create factory
    mesh::CutFactory tFactory(
        aMesh, &tTopology, {},
        mesh::CutAlgorithm::PellikkaGeneralized
    );
 
    // Set terminals
    tFactory.set_terminals(tTerminals, tShellIndices);
 
    // Run
    tFactory.run();
 
    // Results
    message(InfoLevel::Default, "Generated %lu cuts",
            tFactory.cuts().size());
}

Pattern 3: Direct Cohomology Computation

void compute_cohomology_groups(Mesh* aMesh)
{
    // Build and reduce the simplicial complex on the flagged entities
    mesh::SimplicialComplex tComplex(aMesh, true);
    tComplex.coreduce_complexPellikkaGeneralized();
 
    // Compute cohomology (the constructor computes and cleans the generators)
    mesh::Cohomology tCohomology(&tComplex, aMesh);
 
    // Extract H^1 generators
    Cell<Cochain*>& generators = tCohomology.get_Generators()( 1 );
 
    // Visualize
    tCohomology.create_kGeneratorsField(1, aMesh, "H1");
 
    // Analyze generators
    for (uint i = 0; i < generators.size(); ++i) {
        Cochain* gen = generators(i);
        message(InfoLevel::Default, "Generator %u: %lu edges",
                i, gen->get_simplices_map().size());
    }
}

Pattern 4: Algorithm Performance Comparison

void benchmark_algorithms(Mesh* aMesh)
{
    mesh::Topology tTopology(aMesh);
    tTopology.run();
 
    Cell<mesh::Protoshell*> tProtoshells;
 
    CutAlgorithm algorithms[] = {
        CutAlgorithm::PellikkaGeneralized,
        CutAlgorithm::Pellikka,
        CutAlgorithm::CCR,
        CutAlgorithm::BeltedTree
    };
 
    for (auto alg : algorithms) {
        Timer timer;
 
        mesh::CutFactory tFactory(aMesh, &tTopology, tProtoshells, alg);
        tFactory.run();
 
        uint64_t elapsed = timer.stop();
        message(InfoLevel::Default, "Algorithm %d: %lu ms, %lu cuts",
                static_cast<uint>(alg), elapsed, tFactory.cuts().size());
    }
}

Pattern 5: Thin Shell with Terminals

void process_thin_shell_coil(Mesh* aMesh)
{
    // Topology
    mesh::Topology tTopology(aMesh);
    tTopology.run();
 
    // Protoshell configuration
    mesh::Protoshell* tShell = new mesh::Protoshell();
    tShell->add_sideset(101);  // Top surface
    tShell->add_sideset(102);  // Bottom surface
    tShell->add_terminal(201);  // Input terminal
    tShell->add_terminal(202);  // Output terminal
    tShell->add_material("YBCO");
    tShell->add_thickness(1e-6);
 
    Cell<mesh::Protoshell*> tProtoshells = {tShell};
 
    // Terminals
    Cell<Cell<id_t>> tTerminals = {{301, 302}, {401, 402}};
    Cell<id_t> tShellIndices = {0, 0};  // Both on shell 0
 
    // Factory
    mesh::CutFactory tFactory(
        aMesh, &tTopology, tProtoshells,
        mesh::CutAlgorithm::PellikkaGeneralized
    );
 
    tFactory.set_terminals(tTerminals, tShellIndices);
    tFactory.run();
 
    // Cleanup
    delete tShell;
 
    // Use results
    for (mesh::SideSet* cut : tFactory.cuts()) {
        // Apply jump conditions in FEM
    }
}

Troubleshooting

Problem: Non-Manifold Cut Surfaces

Symptom: Error message "Cut surface is non-manifold" or FEM assembly fails.

Cause: Cohomology generator contains "double pockets" or other topological defects.

Solution:

Try different algorithm:
// PellikkaGeneralized may create complex cuts

// Try original Pellikka for cleaner results

CutFactory tFactory(..., CutAlgorithm::Pellikka);
Enable debug output:
tFactory.run();

// save_debug_meshes() is a member of CutProcessor (public,

// cl_CutProcessor.hpp:116), not of CutFactory. CutFactory calls it

// automatically in debug builds (cl_CutFactory.cpp:477-479, compiled out

// under NDEBUG); it writes one file per cut: cut_<index>.vtk

For curve meshes, CutFactory::save_curve_debug_meshes() is public (cl_CutFactory.hpp:184) and writes curve_<id>.vtk. Inspect either family in ParaView.
Inspect the cleanup that actually runs:
- The shipped path is Cohomology::clean_spfa() → remove_cut_pockets() (cl_Cohomology.hpp:135,145), not the metric-threshold filter described under "Manifold Cleanup" — that scheme is a design of record with no implementation, so there are no cycle-density or size thresholds to tune. See the design note there before going looking for them.

Problem: Incorrect Number of Cuts

Symptom: Expected N cuts, got M cuts (M ≠ N).

Cause: Mesh topology doesn't match expected domain connectivity.

Diagnosis:

mesh::Topology tTopology(aMesh);
tTopology.run();
 
message(InfoLevel::Default, "Phi blocks: %lu",
        tTopology.phi_block_ids().size());
message(InfoLevel::Default, "Interfaces: %lu",
        tTopology.interface_ids().size());

Solution:

Check block domain types: Ensure phi/non-phi classification is correct
Verify mesh connectivity: Use mesh viewer to confirm expected topology
Terminal hints: Provide explicit terminals via set_terminals() to guide computation

Problem: Slow Cohomology Computation

Symptom: run() takes minutes/hours for moderate mesh sizes.

Cause: Using suboptimal algorithm (CCR or original Pellikka).

Solution:

// Switch to generalized Pellikka

CutFactory tFactory(..., CutAlgorithm::PellikkaGeneralized);

If still slow:

Profile: Use Profiler to identify hotspot
Mesh quality: Simplify mesh if possible (remove small features)
MPI: Current implementation is serial (future parallelization planned)

Problem: Abstract Nodes Not Created

Symptom: tFactory.abstract_nodes().size() == 0 after run().

Cause: No cuts were generated (topology is simply-connected).

Diagnosis:

message(InfoLevel::Default, "Num cuts: %lu", tFactory.cuts().size());

Solution:

Check mesh: Verify domain is multiply-connected (e.g., toroid, solenoid)
Topology: Ensure phi/non-phi domains are correctly identified
Terminals: Cuts may not be needed if terminals span all connections

Problem: Jump Condition Not Satisfied

Symptom: FEM solution violates Ampere's law (∮ H·dl ≠ I).

Cause: Cut orientation or DOF assignment incorrect.

Diagnosis:

Visualize cut:
tCohomology.create_kGeneratorsField(1, aMesh, "H1_gen");

// Check edge orientations in ParaView
Check abstract node count:
// Should match number of cuts

BELFEM_ERROR(tFactory.abstract_nodes().size() == tFactory.cuts().size(),

"Abstract node count mismatch");

Solution:

Orientation: Use Homology::generators_orientation() to ensure consistent direction
Static condensation: Verify transformation matrix T incorporates [φ] = I correctly
Gauge: Confirm φ is gauged (set to zero at reference point)

Usage Examples

Example 1: Toroidal Magnet

#include "cl_CutFactory.hpp"
#include "cl_Topology.hpp"
 
int main()
{
    // Load toroidal mesh
    Mesh* tMesh = new Mesh("toroid.exo");
 
    // Analyze topology
    mesh::Topology tTopology(tMesh);
    tTopology.run();
 
    // Create factory (no thin shells)
    mesh::CutFactory tFactory(
        tMesh,
        &tTopology,
        {},  // Empty protoshells
        mesh::CutAlgorithm::PellikkaGeneralized
    );
 
    // Generate cuts
    Timer timer;
    tFactory.run();
    uint64_t elapsed = timer.stop();
 
    message(InfoLevel::Default,
            "Generated %lu cuts in %lu ms",
            tFactory.cuts().size(), elapsed);
 
    // Apply to FEM
    for (mesh::SideSet* cut : tFactory.cuts()) {
        cut->set_domain_type(mesh::DomainType::Cut);
    }
 
    // Create DOF manager
    fem::DofManager* tDofManager = new fem::DofManager(tMesh);
    for (mesh::Node* node : tFactory.abstract_nodes()) {
        tDofManager->add_abstract_dof(node, "phi");
    }
 
    // ... FEM assembly and solve ...
 
    delete tDofManager;
    delete tMesh;
    return 0;
}

Example 2: Superconducting Cable

#include "cl_CutFactory.hpp"
#include "cl_Protoshell.hpp"
#include "cl_Topology.hpp"
 
int main()
{
    Mesh* tMesh = new Mesh("cable.h5");
 
    // Topology
    mesh::Topology tTopology(tMesh);
    tTopology.run();
 
    // Protoshell for thin shell approximation
    mesh::Protoshell* tShell = new mesh::Protoshell();
    tShell->add_sideset(101);  // Cable surface
    tShell->add_terminal(201);  // Input
    tShell->add_terminal(202);  // Output
    tShell->add_material("BSCCO");
    tShell->add_thickness(0.2e-3);  // 0.2 mm
 
    Cell<mesh::Protoshell*> tProtoshells = {tShell};
 
    // Terminals
    Cell<Cell<id_t>> tTerminals;
    tTerminals(0) = {1001, 1002, 1003};  // Input terminal nodes
    tTerminals(1) = {2001, 2002, 2003};  // Output terminal nodes
    Cell<id_t> tShellIndices = {0, 0};
 
    // Factory
    mesh::CutFactory tFactory(
        tMesh, &tTopology, tProtoshells,
        mesh::CutAlgorithm::PellikkaGeneralized
    );
    tFactory.set_terminals(tTerminals, tShellIndices);
    tFactory.run();
 
    // Results
    message(InfoLevel::Default, "Cuts: %lu", tFactory.cuts().size());
    message(InfoLevel::Default, "Abstract DOFs: %lu",
            tFactory.abstract_nodes().size());
 
    // Cleanup
    delete tShell;
    delete tMesh;
    return 0;
}

Example 3: Performance Analysis

#include "cl_Cohomology.hpp"
#include "cl_SimplicialComplex.hpp"
 
void analyze_cohomology_performance(Mesh* aMesh)
{
    // Build simplicial complex on the flagged entities of aMesh
    // (the constructor builds the complex itself)
    Timer t1;
    mesh::SimplicialComplex* tComplex = new mesh::SimplicialComplex(aMesh, /*aPeriodicity=*/false);
    uint64_t time_build = t1.stop();
 
    message(InfoLevel::Default, "Complex construction: %lu ms", time_build);
    message(InfoLevel::Default, "  0-simplices: %u", tComplex->number_of_ksimplices(0));
    message(InfoLevel::Default, "  1-simplices: %u", tComplex->number_of_ksimplices(1));
    message(InfoLevel::Default, "  2-simplices: %u", tComplex->number_of_ksimplices(2));
    message(InfoLevel::Default, "  3-simplices: %u", tComplex->number_of_ksimplices(3));
 
    // Reduction
    Timer t2;
    tComplex->reduce_complexPellikkaGeneralized();
    uint64_t time_reduce = t2.stop();
 
    message(InfoLevel::Default, "Reduction: %lu ms", time_reduce);
    message(InfoLevel::Default, "  0-simplices: %u", tComplex->number_of_ksimplices(0));
    message(InfoLevel::Default, "  1-simplices: %u", tComplex->number_of_ksimplices(1));
    message(InfoLevel::Default, "  2-simplices: %u", tComplex->number_of_ksimplices(2));
    message(InfoLevel::Default, "  3-simplices: %u", tComplex->number_of_ksimplices(3));
 
    // Cohomology
    Timer t3;
    mesh::Cohomology tCohomology(tComplex, aMesh);   // computes the groups
    uint64_t time_cohomology = t3.stop();
 
    message(InfoLevel::Default, "Cohomology: %lu ms", time_cohomology);
    message(InfoLevel::Default, "  H^0: %lu", tCohomology.get_Generators()( 0 ).size());
    message(InfoLevel::Default, "  H^1: %lu", tCohomology.get_Generators()( 1 ).size());
    message(InfoLevel::Default, "  H^2: %lu", tCohomology.get_Generators()( 2 ).size());
 
    delete tComplex;
    delete tFlagged;
}

Summary

The homology module provides a complete computational topology toolkit for finite element electromagnetics:

Component	Purpose	Entry Point
CutFactory	Orchestrates cut generation	Main user class
Cohomology	Computes H^k groups	Direct access for advanced use
Homology	Computes H_k groups	Dual to cohomology
SimplicialComplex	Reduction engine	Internal use
Topology	Mesh analysis	Prerequisite for cuts
Protoshell	Thin shell config	Specialized geometries

Recommended Workflow

Load mesh and ensure proper block/sideset labeling
Analyze topology with Topology::run()
Create CutFactory with PellikkaGeneralized algorithm
Set terminals if using current constraints
Run pipeline via CutFactory::run()
Extract results (cuts, abstract nodes)
Integrate with FEM (DOF assignment, jump conditions)

Performance Best Practices

Use PellikkaGeneralized (the production default; no timing comparison is recorded)
Enable debug output only when troubleshooting (overhead ~10%)
Pre-filter mesh to remove unnecessary entities before cohomology computation
For very large meshes (>10^7 elements), consider domain decomposition (future feature)

Module Contracts

Critical Constraints

Mesh Mutation Contract

MPI Serial Limitation

Memory Ownership

Thread Safety

Topological vs FEM Correctness

Topological Guarantees

FEM Guarantees

Mesh Requirements

Scratch-Field Usage

Table of Contents

Quick-Start

Glossary

Common Pitfalls

1. Building the CutFactory Before the Topology Has Run

2. Running Homology After Mesh Partitioning

3. Deleting Cuts or Abstract Nodes

4. Forgetting Edge/Face Creation

5. Mixing Chain and Cochain

6. Not Checking Manifold Status

7. Using Wrong Algorithm for Mesh Size

8. Running CutFactory Twice

9. Ignoring Orphaned Nodes

10. Incorrect Terminal-to-Shell Mapping

Introduction

What is a Cohomology Cut?

Module Architecture

File Organization

Dependencies

CutFactory - Main Entry Point

Purpose

Constructor

Key Methods

Workflow Methods

Result Retrieval

Debugging

Usage Pattern

Cohomology Computation

Purpose

Constructor

Key Methods

Visualization

Usage Pattern

Homology Computation

Purpose

Constructor

Key Methods

Usage Pattern

Cut Algorithm Selection

Available Algorithms

Algorithm Comparison

Implementation Details

Selection Guidance

Topology Analysis

Purpose

Constructor

Key Methods

Usage Pattern

Thin Shell Handling

Purpose

Structure

Key Methods

Usage Pattern

Cut Processing

Purpose

Constructor

Key Methods

Internal Workflow

Manifold Cleanup

Data Structures

Chain

Cochain

CutData

SimplicialComplex

Cut Sideset Creation

InterfaceProcessor

CutSet

Algorithm Classes and Functions

BeltedTree