HaomingLuo's picture
Add DENIM capability extension v4
4452800 verified
|
Raw History Blame Contribute Delete
2.85 kB
# DENIM Capability Extension v4
This increment adds 1,000 complete trajectories to the staged AgentFEM
Material Loading Memory dataset. It targets capability boundaries that were
underrepresented in the frozen 2,660-trajectory release.
## Cohorts
| Cohort | Trajectories | Purpose |
|---|---:|---|
| Long-cycle | 300 | 10, 25, 50 and held-out 100-cycle state evolution |
| Full 3D non-proportional | 300 | Six Voigt components, rotating axes, hydrostatic/deviatoric sequencing and random blocks |
| Structure-derived | 200 | Local histories from four AgentFEM perforated/notched plate unit-load bases |
| Discretization and observation | 200 | 50 paths at 61/121/241/481 states with clean truth plus noisy sparse observations |
One sample is one complete trajectory. Noise and observation masks are stored
as additional fields and are not counted as separate physical samples.
Matched time resolutions share an explicit `group_id` and must remain together
when constructing evaluation protocols.
## Structure-derived fidelity
AgentFEM solves three linear-elastic unit-load cases on each of four 2D
structures. Local DG0 strain bases are superposed under time-dependent load
coefficients and replayed through the hidden nonlinear reference material.
These are genuine finite-element-derived local path shapes, but they do not
include coupled elastoplastic stress redistribution. They are therefore a
bridge between hand-designed material-point paths and global nonlinear FEM,
not a substitute for the latter.
## Main files
- `cohort.h5`: lossless trajectories and provenance fields;
- `index.jsonl` and `index.csv`: viewer-friendly trajectory metadata;
- `manifest.json`: counts, hashes, physical quality gates and AgentFEM commit;
- `structure_bases.npz`: the four groups of AgentFEM local strain bases;
- `configs/t2_denim_capability_v4.json`: frozen design;
- `src/generate_t2_denim_capability_v4.py`: complete generator;
- `src/analyze_t2_denim_capability_v4.py`: paired-resolution and observation audit.
## Fields
Every trajectory contains tensor-shear Voigt arrays in the order
`xx, yy, zz, xy, yz, xz`: total strain, stress, plastic strain, two visible
memory channels, three reference memory channels, PEEQ, plastic increment,
isotropic radius and normalized history coordinate. Observation trajectories
also contain `observed_strain`, `observed_stress_pa`, and `observation_mask`.
Structure-derived trajectories additionally retain the source coordinate,
cell index, three local unit-load strain bases and load-mode coefficients.
## Scope
The data are synthetic, small-strain and rate-independent. The reference is a
fixed three-memory Chaboche-type material with tabulated isotropic hardening.
The release is not experimental calibration, fatigue-life data, damage data,
finite-strain plasticity, or a certified material card.