PAPER_ROADMAP.md

# Paper roadmap: learned constitutive memory with controlled generalization

## Recommended first paper

**Working title:** *A verified benchmark for history-dependent neural
constitutive modeling under in-distribution and out-of-distribution loading*

The central question should be: **when a learned constitutive surrogate is
accurate on held-out trajectories, does it also preserve material memory and
physical admissibility under new parameters, new loading paths, new temporal
resolution and structural deployment?**

This is stronger than a simple “GRU outperforms MLP” result. The current v1
dataset already establishes the need for history. The paper should investigate
which representations and models generalize reliably, where they fail and how
physics-based checks alter the conclusion.

## Evidence already available

- A deterministic, balanced, quality-gated 1,008-trajectory benchmark.
- J2 linear-isotropic and Chaboche combined-hardening labels.
- Six proportional loading families and leakage-safe trajectory splits.
- Analytical J2 checks, yield-surface residuals and 121-to-241-state audits.
- Pointwise MLP control: 244.54 MPa RMSE and 0.4316 test R2.
- History-aware GRU baseline: 16.52 MPa RMSE and 0.9974 test R2.
- Complete generation, reading, training and testing code.

## Required additions before journal submission

### A. True multiaxial non-proportional extension

Add tensor loading paths whose deviatoric direction changes with time, for
example tension-shear phase shifts, rotating principal strain directions,
piecewise changes of loading direction and mixed axial/shear reversals. Keep
proportional v1 untouched and publish this as a versioned extension.

Minimum evidence:

- tensor-valued strain and stress histories;
- full backstress tensors and internal variables;
- independent verification on selected paths;
- time-step refinement and failure-ledger reporting;
- balanced path and parameter coverage.

### B. Generalization splits that answer scientific questions

Evaluate at least four protocols:

1. **ID trajectory split:** current leakage-safe v1 split.
2. **Path OOD:** withhold complete loading families during training.
3. **Parameter OOD:** withhold bounded parameter regions, including controlled
   interpolation and extrapolation tests.
4. **Resolution OOD:** train at one temporal discretization and test at finer,
   coarser and irregular sampling.

For multiaxial data, add a direction-change or phase-angle OOD protocol. Freeze
all splits before final model comparison.

### C. Model comparison and ablation

Use compact, reproducible models rather than an indiscriminate architecture
survey:

- history-blind pointwise MLP control;
- incremental state MLP with explicitly supplied internal state;
- GRU and LSTM;
- one causal transformer, temporal convolution or state-space model;
- one physics-regularized variant of the strongest compact model.

Ablate material parameters, history length, internal-state supervision,
physical penalties and training-set size. Match parameter counts or report them
clearly. Include training and inference cost.

### D. Physics-aware evaluation

Report conventional stress errors together with:

- peak and reversal stress error;
- yield-onset error;
- hysteresis-loop area error;
- ratcheting or mean-stress evolution error where applicable;
- yield-surface residual;
- plastic incompressibility error;
- nondecreasing equivalent-plastic-strain violations;
- long-rollout drift and stability;
- uncertainty or failure detection outside the training domain.

Do not label raw stress work as thermodynamic dissipation until recoverable
backstress energy is represented consistently.

### E. Structural finite-element deployment

Embed the selected learned material update in at least one small structural
problem, such as a cyclically loaded coupon and a notched coupon. Compare with
the native AgentFEM constitutive update for:

- force-displacement loops;
- local stress and plastic-strain histories;
- convergence and stability;
- accumulated error over cycles;
- wall-clock time and speed-up.

This step demonstrates that a material-point score transfers—or does not
transfer—to a finite-element simulation.

### F. Reproducible release

Publish the added data, frozen splits, training configurations, seeds, model
weights, evaluation scripts, environment files and selected FE cases. Include
an independent download-and-run audit in a clean environment.

## Suggested paper figures

1. Dataset and quality-control workflow.
2. Proportional and non-proportional loading-path atlas.
3. Representative reference and predicted hysteresis loops.
4. ID versus path/parameter/resolution OOD performance.
5. Physics-violation comparison across models.
6. Data-size and model-ablation curves.
7. Structural coupon force-displacement and local-field comparison.
8. Accuracy-cost-stability trade-off.

## Publication assessment

The current v1 release alone supports a useful dataset or benchmark note. If
sections A–F are completed rigorously, the work can support a substantive
computational mechanics or scientific-machine-learning article. A strong
specialist journal is realistic. A top-tier mechanics journal would additionally
require a clear methodological advance—such as a materially superior
physics-consistent state representation—not only more trajectories and model
comparisons.