REPRODUCE.md

# Reproducing AgentFEM Material Loading Memory v1

## Reproducibility levels

The release supports three distinct levels. Do not confuse them.

1. **Data verification** checks the eight HDF5 shard hashes, schema, trajectory
   IDs, splits, model labels, path-family counts and finite values. It requires
   only Python, NumPy and h5py.
2. **Baseline reproduction** retrains the published MLP and GRU from the frozen
   trajectory splits. CPU training is sufficient.
3. **Physics regeneration** recreates all trajectories with the exact AgentFEM
   source revision, repackages the shards and reruns the numerical audit.

## Obtain the release

Using the Hugging Face CLI:

```bash
hf download HaomingLuo/AgentFEM-Material-Loading-Memory \
  --repo-type dataset \
  --local-dir AgentFEM-Material-Loading-Memory
cd AgentFEM-Material-Loading-Memory
```

## Verify the downloaded data

```bash
conda env create -f environment-use.yml
conda activate agentfem-t2-use
bash reproduce_t2_v1.sh verify
```

Expected summary:

- 1,008 unique trajectories;
- 768/120/120 train/validation/test trajectories;
- 504 J2 and 504 Chaboche trajectories;
- 168 trajectories in each of six path families;
- eight shard hashes matching `manifest.json`.

## Reproduce the baselines

```bash
bash reproduce_t2_v1.sh baseline
```

The script trains with the published defaults: 40 MLP epochs and 60 GRU
epochs. Small numerical variation across PyTorch versions and hardware is
normal. Compare with `artifacts/t2_material_loading_memory_v1/baseline_metrics.json`,
not by requiring bitwise-identical neural-network weights.

## Regenerate the physics data

The frozen physics environment uses Python 3.11, FEniCSx/DOLFINx 0.11.0 and
AgentFEM commit `058faecc05aeda143d014fd229401003a9258bbb`.

```bash
conda env create -f environment-reproduce.yml
conda activate agentfem-t2-reproduce
bash reproduce_t2_v1.sh design
bash reproduce_t2_v1.sh full
```

`design` checks the deterministic 1,008-case Sobol design without solving.
`full` performs the following steps:

1. regenerate or resume all 1,008 material-point trajectories;
2. build eight HDF5 shards;
3. run analytical, constitutive and time-resolution audits;
4. retrain both reference baselines;
5. run the published automated tests.

The generator is restartable. For controlled parallel work, different workers
may run non-overlapping ranges before one worker packages and audits:

```bash
python src/t2_material_loading_memory_v1.py --range 0 252
python src/t2_material_loading_memory_v1.py --range 252 504
python src/t2_material_loading_memory_v1.py --range 504 756
python src/t2_material_loading_memory_v1.py --range 756 1008
python src/t2_material_loading_memory_v1.py --package
python src/t2_material_loading_memory_v1.py --audit
```

## Exact source fallback

If the VCS entry in `environment-reproduce.yml` cannot install AgentFEM,
install it explicitly after creating the environment:

```bash
git clone https://github.com/haoming-luo/agentfem.git
cd agentfem
git checkout 058faecc05aeda143d014fd229401003a9258bbb
python -m pip install .
cd ..
```

Confirm before regeneration:

```bash
python -c "import agentfem; print(agentfem.__version__)"
git -C agentfem rev-parse HEAD
```

## Reproducibility contract for extensions

- A sample is one complete trajectory; frames must not cross data splits.
- Keep SI units and record all material parameters and load-path controls.
- Preserve v1 as an immutable benchmark. Publish extensions under a new
  dataset version and document migration rules.
- Failed or non-converged cases must be retained in a failure ledger rather
  than silently removed.
- Split generation must be deterministic and grouped by trajectory and source
  design.
- Report both predictive errors and physical-consistency errors.