# Start here: AgentFEM Material Loading Memory v1

This repository is a research handoff, not only a file archive. It contains the
published dataset, the exact generation configuration, reusable readers,
reference baselines, numerical tests, quality evidence and a concrete route for
extending the work into a research article.

## Choose your route

### 1. Use the data or train a new model

You do **not** need AgentFEM. Create the lightweight environment and verify the
download:

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

Start with `src/load_t2_material_loading_memory.py`. The current HDF5 data can
be used directly for GRU, LSTM, transformer, state-space or operator-learning
experiments.

### 2. Reproduce the published baseline

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

This verifies the data and retrains the pointwise MLP and history-aware GRU.
The published reference metrics are in
`artifacts/t2_material_loading_memory_v1/baseline_metrics.json`.

### 3. Regenerate the physics data

Use `environment-reproduce.yml`, which installs FEniCSx and the exact AgentFEM
commit used for v1:

```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
```

Full regeneration is deterministic but rewrites generated shards and baseline
artifacts in the working copy. See `REPRODUCE.md` before running it.

## What is already established

- 1,008 complete trajectories, not 121,968 independent frames.
- Two explicitly labelled constitutive models: J2 linear isotropic hardening
  and Chaboche combined hardening.
- Six balanced proportional loading-path families.
- Complete-trajectory train/validation/test splits with no frame leakage.
- Numerical consistency, analytical, refinement and checksum evidence.
- A strong history-aware baseline and a deliberately history-blind control.

## What remains for a strong paper

The current release is a verified foundation, not the full paper claim. The
recommended first paper studies generalization of learned path-dependent
constitutive models. Its priority additions are true non-proportional
multiaxial paths, physically meaningful out-of-distribution splits,
physics-aware metrics, stronger sequence-model comparisons and deployment in a
small structural finite-element example. See `PAPER_ROADMAP.md`.

## Source of truth and citation

- Dataset: <https://huggingface.co/datasets/HaomingLuo/AgentFEM-Material-Loading-Memory>
- AgentFEM: <https://github.com/haoming-luo/agentfem>
- AgentFEM commit used for v1:
  `058faecc05aeda143d014fd229401003a9258bbb`
- Dataset license: CC BY 4.0
- Included source-code license: Apache-2.0

When extending the dataset, preserve trajectory-level grouping, units, model
labels, random seeds, quality failures and the original v1 test set. Record a
new schema or dataset version whenever fields or physical scope change.