Skip to content

AgentFEM 0.3.0

AgentFEM 0.3.0 is the first release built on a stable scientific middle layer between humans or AI agents and the FEniCSx/PETSc numerical runtime.

One public workflow

The recommended language is:

Study -> Model -> scientific assets -> model.step(...) -> SimulationResult

Model remains a readable engineering facade. Constitutive owns material updates, State owns committed and trial history, Operator owns mathematical contributions, Procedure owns numerical evolution, Backend owns finite-element execution, and Result/Verification owns evidence and acceptance. Architecture tests keep these responsibilities from silently collapsing back into one numerical god object.

Evidence-bound execution

The same execution lifecycle now carries solver options, output, history, progress and checkpoint policy. Structured results retain convergence, balances, failures, provenance, scientific quantities and artifact integrity. Capability maturity remains bounded by executable Golden, failure, MPI or external evidence rather than by the existence of an API name.

Open extension seam

AgentFEM-Learning demonstrates that a separately installed package can register neural-field and finite-domain XDEM providers without modifying the core. The extension is discovered through the same provider boundary and returns the same SimulationResult contract used by deterministic workflows.

Human and agent acceptance

The release candidate passed clean installed-wheel acceptance on hosted Linux and macOS. A genuinely fresh Codex session then consumed the immutable wheel, inspected the runtime, constructed and refined a plane-strain cantilever, verified force and work balance plus artifact integrity, and explained the model assumptions without human correction. These are behavioral acceptance conditions, not deterministic CI substitutes.

Scientific foundation in this release

The 0.3 foundation includes reviewed Abaqus migration, solver-neutral user materials, transient and nonlinear State transactions, finite-strain J2 RVE workflows, mixed-mode LEFM extraction, cohesive and creep foundations, parameter campaigns, scientific datasets and learning interfaces. Each keeps its declared maturity and applicability boundary.

Scope

AgentFEM 0.3.0 does not claim universal element, material, contact or fracture coverage. Native Windows, general contact, arbitrary-path fracture, monolithic thermo-mechanical coupling, and automatic PINN or neural-operator training for arbitrary meshes remain outside the release claim. WSL2 remains the recommended Windows route and a separately tracked platform target.