Skip to content

AgentFEM 0.3.8

AgentFEM 0.3.8 is the architectural-foundation release before 0.4. It strengthens the boundaries between Model, Operator, Procedure, State, Backend, and Result/Verification instead of adding another disconnected solver family.

Foundation gates

The release closes six declared foundation gates:

  1. one mesh--element--function-space compatibility contract used by inspection, validation, lowering, result identity, and checkpoint identity;
  2. mesh-quality evidence and early rejection of unsupported topology or formulation combinations;
  3. typed time inputs that distinguish right-hand-side, operator, state, and output-only effects before a Procedure reuses or rebuilds an operator;
  4. provider-owned reaction, accepted-path work, and energy evidence for the declared MPC, weak-constraint, and bounded contact routes;
  5. a typed MPI result, state-identity, and checkpoint lifecycle with verified cross-rank restoration where portability is declared;
  6. an independently installed extension that adds behavior through the public provider boundary without modifying AgentFEM core.

Contact and nonlinear state

The first contact routes remain intentionally bounded. The nonlinear-static route covers small-strain solid mechanics, one analytical rigid plane and frictionless penalty contact, with potential, residual, tangent, nodal reaction, resultant, penetration diagnostics, accepted-path work semantics, and conservative energy evidence.

The bounded explicit route now accepts one proportional prescribed rigid translation/rotation schedule. It evaluates the moving surface at the predicted displacement, publishes MPI-global force and moment, integrates only accepted force--translation/moment--rotation stations, rolls failed increments back, and restores the accepted work path from transient checkpoints. Serial restart and two-rank canonical-State tests protect that lifecycle. The bounded DOLFINx route can now derive a conservative contact spectral contribution from its trace, both penalty channels, Coulomb pressure-cap coupling, and the actual lumped mass for fixed-normal or piecewise-planar projectors; the Procedure adds it to the body/material spectral bound before selecting the time increment. Bounded rigid pairs compose through the residual chain: accepted time, rollback, per-pair force/moment, summed work/potential, and nested restart propagate without a forming-specific solver; a mixed analytical-plus-triangulated pair test uses that route without a two-tool special case. A separate reviewed path couples the same Procedure to serial and routed distributed triangle BVHs; stable slave-point identity is preserved while the closest rigid facet changes during finite sliding, and pure tangential tool translation produces no spurious normal contact work. The recommended construction binds surface, schedule, reference point, and rank-independent scientific identity in one immutable RigidBody; serial or distributed search remains a separate backend object rather than becoming part of the body's identity. RigidContactPair binds the slave boundary, body, and scalar local law as one inspectable Model asset while leaving search and time-step stability with the Backend and Procedure. Pair and body identities are verified before checkpoint restoration. Accepted transient events expose compact contact activity, maximum penetration, force norm, prescribed-motion work, and interval power through the existing bounded/throttled progress lifecycle; no per-point history is accumulated implicitly.

The fixed public Abaqus/Explicit finite-sliding protocol now has a complete AgentFEM solid bridge rather than only component tests. It uses the published material, 500-unit normal load, friction coefficient 0.3, and 0.1-unit slide; the ordinary Procedure reaches frictionless preload equilibrium, transfers the accepted state atomically, and then closes normal force, the Coulomb cap, action--reaction, friction dissipation, energy, and serial/two-rank endpoint checks. The bounded tessellated tool also records real stable-facet crossings without changing slave-point identity. Because the source discretization is B31 and the bridge is a tetrahedral CG1 solid, the evidence is labelled a protocol bridge rather than an elementwise external reproduction.

The bridge additionally separates time and space refinement. It first halves both stage time increments without changing physical duration, then refines the in-plane mesh while retaining those time increments and durations. The certificate requires stable friction dissipation, persistent facet crossings, and non-growing energy error; it deliberately does not infer an observed order from a two-level study.

General multi-tool candidate routing, arbitrary surface topology, free rigid-body dynamics, deformable-to-deformable contact, and forming are not inferred from this evidence. They remain explicit future capabilities.

The 0.4 foundation work now also defines a solver-neutral accepted/trial closest-point State. It keeps stable 64-bit contact-point and surface-facet identity, permits facet changes during sliding, rejects contact-pair identity changes, and checkpoints only accepted projection boundaries. The explicit contact residual now gathers those accepted records by global slave-point ID and restores them onto the current MPI partition, including master facet and barycentric coordinates. Search remains geometry infrastructure: restored records preserve audit continuity, while the next residual evaluation still performs exact projection and never assembles force from a stale gap or normal. No general implicit Newton update policy is claimed by this state contract.

A separate projection lifecycle now owns every-evaluation calls to analytical or reviewed BVH projectors, keeps search diagnostics out of State, and binds projection commit/rollback to the accepted increment. Distributed projectors reach collective post-search acceptance, preventing one rank from entering assembly after another rank has rejected invalid projection evidence. This is still a lifecycle foundation, not a general contact Operator.

The same foundation now includes a solver-neutral local frictionless penalty law. It consumes any reviewed SurfaceProjection, returns penetration, pressure, conservative potential density, structural residual traction, and the conjugate rigid-surface traction, and rejects invalid search evidence by default. The response is deliberately labelled pointwise and unintegrated: trace interpolation, quadrature, global assembly, and geometry-consistent linearization remain required before AgentFEM can claim a general implicit contact Operator.

A backend-neutral ContactTrace now defines the next hand-off without hard-coding one element family. It binds stable point IDs to node interpolation, positive quadrature weights, and an explicit reference/current measure, then integrates a matching projection record and point response into nodal residual, penalty potential, contact resultant, and optional tool moment. It rejects identity, coordinate, or response mismatches before assembly. This is the backend-neutral reference trace contract; consistent implicit linearization remains outstanding.

The first DOLFINx trace adapter now closes that hand-off for one bounded route: tagged owned exterior triangles on first-order tetrahedral geometry and a continuous blocked three-component CG1 displacement space. It uses a positive three-point reference-area rule, derives stable point identity from the partition-independent facet IDs, synchronizes displacement ghosts before interpolation, supports empty local shards, and rejects unsupported spaces collectively. General/high-order topology and a consistent implicit contact linearization are still not claimed.

Time and restart semantics

Time-dependent inputs now declare whether they affect only the right-hand side, the operator, constitutive state, or output. Linear implicit dynamics and implicit-Euler heat transfer use this information to select safe reuse; nonlinear heat transfer continues to assemble per step. Transient checkpoint schema v5 binds the complete input plan and rejects changed or unidentifiable callbacks before mutating a restored field.

The public checkpoint capability contract separates payload scope, save boundary, atomicity, scientific identity, and MPI rank-count portability. Supported transient, nonlinear, inelastic, viscoelastic, harmonic, and cyclic owners expose that contract through model.step() and SimulationResult.

Mesh and extension evidence

Reviewed higher-order triangle, quadrilateral, tetrahedral, and hexahedral routes now preserve coordinate-element identity through import, quality audit, patch testing, output, and supported MPI paths. The six-node wedge route is also promoted for its declared path. Conditional or unsupported topology is reported before assembly rather than failing inside a low-level reader.

The reference external material is built and installed as a separate wheel. The extension acceptance proves provider discovery, ordinary model execution, structured results, verification, and unchanged core-package hashes.

Release CI also creates a fresh-agent trial bundle bound to the exact candidate wheel, source commit, task, review contract, required command sequence, and required outputs. This prevents an earlier successful trial from being reused as evidence for a different candidate.

Upgrade

Install or upgrade with conda-forge after its feedstock update is available:

mamba install -c conda-forge agentfem=0.3.8

The Python distribution can also be upgraded after PyPI publication:

python -m pip install --upgrade agentfem==0.3.8

Existing projects keep the public Study -> Model -> scientific assets -> model.step(...) -> SimulationResult workflow. Checkpoint identity is stricter by design; incompatible or stale archives fail before partial restoration.

Explicit boundaries

AgentFEM 0.3.8 does not claim universal finite-element coverage, native Windows solver support, general contact, forming-capable shells, arbitrary-path fracture, universal nonlinear-material validation, or monolithic multiphysics. Use agentfem capabilities for the exact installed maturity and evidence.