# DENIM incomplete-physics closure

DENIM is the **Discrete-Energy Neural Internal-variable Model** developed in
the AgentFEM material-loading-memory study.

This increment contains 128 complete multiaxial trajectories, each with 241
ordered states. Training uses five path families (80 trajectories), validation
uses rotating-principal loading (16), and test fully withholds out-of-phase
Lissajous and random-direction-block loading (32).

The reference is intentionally richer than the learned model:

- reference: three kinematic memories plus tabulated isotropic hardening;
- DENIM: two tensor memories plus a learned monotone saturating hardening law;
- known to DENIM: isotropic elasticity, J2 geometry and initial yield stress;
- hidden: reference hardening table, recovery parameters and update equations.

Start with:

- `data/t2_graybox_closure_v1/manifest.json`
- `artifacts/t2_graybox_closure_v1/model_metrics.json`
- `artifacts/t2_graybox_closure_v1/deployment_validation.json`
- `docs/T2_GRAYBOX_CLOSURE_RESULTS.md`

The HDF5 groups store strain, stress, plastic strain, PEEQ, plastic increments,
the three reference memories and the two-channel closure state. SI units and
Voigt order `xx, yy, zz, xy, yz, xz` with tensor shear are used throughout.

Data are CC BY 4.0. Included code uses the repository code license.