HaomingLuo commited on
Commit
f1564a0
·
verified ·
1 Parent(s): 9b466fd

Add multiaxial OOD v2 data, six neural models, and FE validation

Browse files

Adds 1,024 verified multiaxial J2/Chaboche trajectories, ID/path/parameter OOD splits, six model families including a physics-embedded neural stress integrator, structural FE gates, reproducibility code, checksums, and research notes. Preserves the published v1 release.

This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. KNOWN_LIMITATIONS.md +47 -0
  2. PAPER_ROADMAP.md +131 -0
  3. README.md +70 -88
  4. REPRODUCE.md +116 -0
  5. SHA256SUMS +93 -1
  6. START_HERE.md +80 -0
  7. T2_V2_REPRODUCE.md +54 -0
  8. T2_V2_START_HERE.md +42 -0
  9. artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md +20 -0
  10. artifacts/t2_multiaxial_ood_v2/model_metrics.json +290 -0
  11. artifacts/t2_multiaxial_ood_v2/model_protocol_comparison.png +3 -0
  12. artifacts/t2_multiaxial_ood_v2/multiaxial_hysteresis_preview.png +3 -0
  13. artifacts/t2_multiaxial_ood_v2/physics_integrator_summary.json +78 -0
  14. artifacts/t2_multiaxial_ood_v2/quality.json +228 -0
  15. artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/notched_bar_comparison.png +3 -0
  16. artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/notched_bar_results.npz +3 -0
  17. artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/structural_validation.json +51 -0
  18. artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/notched_bar_comparison.png +3 -0
  19. artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/notched_bar_results.npz +3 -0
  20. artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/structural_validation.json +41 -0
  21. artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/notched_bar_comparison.png +3 -0
  22. artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/notched_bar_results.npz +3 -0
  23. artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/structural_validation.json +45 -0
  24. artifacts/t2_multiaxial_ood_v2/training_summary.json +18 -0
  25. configs/t2_multiaxial_ood_v2.json +65 -0
  26. data/t2_multiaxial_ood_v2/README.md +114 -0
  27. data/t2_multiaxial_ood_v2/design.jsonl +0 -0
  28. data/t2_multiaxial_ood_v2/index.jsonl +0 -0
  29. data/t2_multiaxial_ood_v2/manifest.json +84 -0
  30. data/t2_multiaxial_ood_v2/shards/part-00000.h5 +3 -0
  31. data/t2_multiaxial_ood_v2/shards/part-00001.h5 +3 -0
  32. data/t2_multiaxial_ood_v2/shards/part-00002.h5 +3 -0
  33. data/t2_multiaxial_ood_v2/shards/part-00003.h5 +3 -0
  34. data/t2_multiaxial_ood_v2/shards/part-00004.h5 +3 -0
  35. data/t2_multiaxial_ood_v2/shards/part-00005.h5 +3 -0
  36. data/t2_multiaxial_ood_v2/shards/part-00006.h5 +3 -0
  37. data/t2_multiaxial_ood_v2/shards/part-00007.h5 +3 -0
  38. docs/T2_MULTIAXIAL_LITERATURE.md +67 -0
  39. docs/T2_MULTIAXIAL_RESEARCH_MEMO.md +105 -0
  40. environment-reproduce.yml +15 -0
  41. environment-use.yml +10 -0
  42. models/t2_multiaxial_ood_v2/README.md +70 -0
  43. models/t2_multiaxial_ood_v2/id_causal_tcn.pt +3 -0
  44. models/t2_multiaxial_ood_v2/id_causal_tcn_history.json +177 -0
  45. models/t2_multiaxial_ood_v2/id_gru.pt +3 -0
  46. models/t2_multiaxial_ood_v2/id_gru_history.json +177 -0
  47. models/t2_multiaxial_ood_v2/id_lstm.pt +3 -0
  48. models/t2_multiaxial_ood_v2/id_lstm_history.json +177 -0
  49. models/t2_multiaxial_ood_v2/id_normalization.json +93 -0
  50. models/t2_multiaxial_ood_v2/id_physics_integrator_nn.pt +3 -0
KNOWN_LIMITATIONS.md ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Known limitations and safe claims
2
+
3
+ ## What v1 represents
4
+
5
+ - Synthetic, three-dimensional, small-strain material-point histories.
6
+ - Prescribed proportional deviatoric strain paths.
7
+ - J2 plasticity with linear isotropic hardening.
8
+ - Chaboche combined hardening, explicitly marked as an experimental AgentFEM
9
+ capability.
10
+ - A benchmark for path dependence, sequence learning and reproducible
11
+ constitutive-surrogate studies.
12
+
13
+ ## What v1 does not represent
14
+
15
+ - Structural finite-element field data.
16
+ - Experimental calibration of a named alloy.
17
+ - Fatigue life, damage initiation or crack propagation.
18
+ - Temperature dependence, creep, viscoplasticity or finite strain.
19
+ - True non-proportional multiaxial loading with rotating stress/strain
20
+ directions.
21
+ - Validation against every commercial constitutive implementation.
22
+
23
+ ## Important interpretation limits
24
+
25
+ - High random-split accuracy does not establish path or parameter
26
+ extrapolation.
27
+ - The pointwise MLP is a history-blind scientific control, not a competitive
28
+ constitutive architecture.
29
+ - The GRU result establishes learnability on the frozen v1 distribution; it
30
+ does not by itself establish stable long-cycle or FE deployment.
31
+ - Material-point inference speed does not equal end-to-end structural speed-up.
32
+ - Raw stress work on plastic strain is retained as a diagnostic. For the
33
+ Chaboche model it must not be called thermodynamic dissipation without a
34
+ consistent split of stored and dissipated backstress energy.
35
+ - Neural-network metrics may vary slightly across hardware and library
36
+ versions even when the physics data are deterministic.
37
+
38
+ ## Changes that require a new version
39
+
40
+ - adding tensor loading paths or new constitutive models;
41
+ - changing physical parameter ranges or path-generation rules;
42
+ - changing stored variables, units or schema semantics;
43
+ - revising trajectory grouping or benchmark splits;
44
+ - correcting any physical or numerical defect in a released trajectory.
45
+
46
+ Extensions should retain v1 unchanged, publish migration notes and report
47
+ results separately on the original v1 test set and the new OOD test sets.
PAPER_ROADMAP.md ADDED
@@ -0,0 +1,131 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Paper roadmap: learned constitutive memory with controlled generalization
2
+
3
+ ## Recommended first paper
4
+
5
+ **Working title:** *A verified benchmark for history-dependent neural
6
+ constitutive modeling under in-distribution and out-of-distribution loading*
7
+
8
+ The central question should be: **when a learned constitutive surrogate is
9
+ accurate on held-out trajectories, does it also preserve material memory and
10
+ physical admissibility under new parameters, new loading paths, new temporal
11
+ resolution and structural deployment?**
12
+
13
+ This is stronger than a simple “GRU outperforms MLP” result. The current v1
14
+ dataset already establishes the need for history. The paper should investigate
15
+ which representations and models generalize reliably, where they fail and how
16
+ physics-based checks alter the conclusion.
17
+
18
+ ## Evidence already available
19
+
20
+ - A deterministic, balanced, quality-gated 1,008-trajectory benchmark.
21
+ - J2 linear-isotropic and Chaboche combined-hardening labels.
22
+ - Six proportional loading families and leakage-safe trajectory splits.
23
+ - Analytical J2 checks, yield-surface residuals and 121-to-241-state audits.
24
+ - Pointwise MLP control: 244.54 MPa RMSE and 0.4316 test R2.
25
+ - History-aware GRU baseline: 16.52 MPa RMSE and 0.9974 test R2.
26
+ - Complete generation, reading, training and testing code.
27
+
28
+ ## Required additions before journal submission
29
+
30
+ ### A. True multiaxial non-proportional extension
31
+
32
+ Add tensor loading paths whose deviatoric direction changes with time, for
33
+ example tension-shear phase shifts, rotating principal strain directions,
34
+ piecewise changes of loading direction and mixed axial/shear reversals. Keep
35
+ proportional v1 untouched and publish this as a versioned extension.
36
+
37
+ Minimum evidence:
38
+
39
+ - tensor-valued strain and stress histories;
40
+ - full backstress tensors and internal variables;
41
+ - independent verification on selected paths;
42
+ - time-step refinement and failure-ledger reporting;
43
+ - balanced path and parameter coverage.
44
+
45
+ ### B. Generalization splits that answer scientific questions
46
+
47
+ Evaluate at least four protocols:
48
+
49
+ 1. **ID trajectory split:** current leakage-safe v1 split.
50
+ 2. **Path OOD:** withhold complete loading families during training.
51
+ 3. **Parameter OOD:** withhold bounded parameter regions, including controlled
52
+ interpolation and extrapolation tests.
53
+ 4. **Resolution OOD:** train at one temporal discretization and test at finer,
54
+ coarser and irregular sampling.
55
+
56
+ For multiaxial data, add a direction-change or phase-angle OOD protocol. Freeze
57
+ all splits before final model comparison.
58
+
59
+ ### C. Model comparison and ablation
60
+
61
+ Use compact, reproducible models rather than an indiscriminate architecture
62
+ survey:
63
+
64
+ - history-blind pointwise MLP control;
65
+ - incremental state MLP with explicitly supplied internal state;
66
+ - GRU and LSTM;
67
+ - one causal transformer, temporal convolution or state-space model;
68
+ - one physics-regularized variant of the strongest compact model.
69
+
70
+ Ablate material parameters, history length, internal-state supervision,
71
+ physical penalties and training-set size. Match parameter counts or report them
72
+ clearly. Include training and inference cost.
73
+
74
+ ### D. Physics-aware evaluation
75
+
76
+ Report conventional stress errors together with:
77
+
78
+ - peak and reversal stress error;
79
+ - yield-onset error;
80
+ - hysteresis-loop area error;
81
+ - ratcheting or mean-stress evolution error where applicable;
82
+ - yield-surface residual;
83
+ - plastic incompressibility error;
84
+ - nondecreasing equivalent-plastic-strain violations;
85
+ - long-rollout drift and stability;
86
+ - uncertainty or failure detection outside the training domain.
87
+
88
+ Do not label raw stress work as thermodynamic dissipation until recoverable
89
+ backstress energy is represented consistently.
90
+
91
+ ### E. Structural finite-element deployment
92
+
93
+ Embed the selected learned material update in at least one small structural
94
+ problem, such as a cyclically loaded coupon and a notched coupon. Compare with
95
+ the native AgentFEM constitutive update for:
96
+
97
+ - force-displacement loops;
98
+ - local stress and plastic-strain histories;
99
+ - convergence and stability;
100
+ - accumulated error over cycles;
101
+ - wall-clock time and speed-up.
102
+
103
+ This step demonstrates that a material-point score transfers—or does not
104
+ transfer—to a finite-element simulation.
105
+
106
+ ### F. Reproducible release
107
+
108
+ Publish the added data, frozen splits, training configurations, seeds, model
109
+ weights, evaluation scripts, environment files and selected FE cases. Include
110
+ an independent download-and-run audit in a clean environment.
111
+
112
+ ## Suggested paper figures
113
+
114
+ 1. Dataset and quality-control workflow.
115
+ 2. Proportional and non-proportional loading-path atlas.
116
+ 3. Representative reference and predicted hysteresis loops.
117
+ 4. ID versus path/parameter/resolution OOD performance.
118
+ 5. Physics-violation comparison across models.
119
+ 6. Data-size and model-ablation curves.
120
+ 7. Structural coupon force-displacement and local-field comparison.
121
+ 8. Accuracy-cost-stability trade-off.
122
+
123
+ ## Publication assessment
124
+
125
+ The current v1 release alone supports a useful dataset or benchmark note. If
126
+ sections A–F are completed rigorously, the work can support a substantive
127
+ computational mechanics or scientific-machine-learning article. A strong
128
+ specialist journal is realistic. A top-tier mechanics journal would additionally
129
+ require a clear methodological advance—such as a materially superior
130
+ physics-consistent state representation—not only more trajectories and model
131
+ comparisons.
README.md CHANGED
@@ -4,106 +4,88 @@ task_categories:
4
  - time-series-forecasting
5
  - other
6
  tags:
 
 
7
  - constitutive-modeling
8
- - computational-mechanics
9
- - scientific-machine-learning
10
- - hysteresis
11
- - agentfem
12
  pretty_name: AgentFEM Material Loading Memory
13
  size_categories:
14
  - 1K<n<10K
15
  ---
16
 
17
- # AgentFEM Material Loading Memory v1
18
 
19
- This dataset contains 1,008 verified synthetic material-point trajectories for
20
- path-dependent small-strain plasticity. It is designed for sequence models,
21
- constitutive surrogates, loading-path generalization, material-model discovery
22
- and reproducible scientific-ML studies.
23
 
24
- ![Loading-path and hysteresis preview](artifacts/t2_material_loading_memory_v1/hysteresis_preview.png)
25
 
26
- ## Dataset summary
 
 
 
 
27
 
28
- - 504 J2 linear-isotropic-hardening trajectories.
29
- - 504 Chaboche combined-hardening trajectories.
30
- - Six balanced loading families: monotonic tension, unload/reload,
31
- tension/compression, symmetric cycling, mean-shifted cycling and variable
32
- amplitude.
33
- - 121 ordered states per trajectory; frames are not counted as independent
34
- samples.
35
- - Trajectory-level split: 768 train, 120 validation and 120 test cases.
36
- - Eight HDF5 shards with 126 independent trajectories per shard.
37
- - SI units, deterministic Sobol parameter design and stable case IDs.
38
 
39
- Every HDF5 group stores prescribed strain, stress, plastic strain, equivalent
40
- plastic strain, plastic multiplier increments, elastic/plastic flags, trial and
41
- corrected yield information, Chaboche backstress state, raw work diagnostics,
42
- material parameters and split metadata.
43
 
44
- ## Verification
 
 
 
 
 
 
 
45
 
46
- All 1,008 trajectories passed finite-value, initial-state, plastic
47
- incompressibility, nondecreasing-PEEQ, plastic-excitation and yield-surface
48
- checks. The maximum yield-surface relative residual was `2.166e-11`; the
49
- maximum J2 monotonic analytical relative error was `4.714e-16`. Twenty-four
50
- representative 121-to-241-state refinement audits produced maximum stress and
51
- PEEQ changes of 0.387% and 0.129%, respectively. Every HDF5 shard is recorded
52
- with a SHA-256 digest in `manifest.json`.
53
 
54
- The complete evidence is in
55
- `artifacts/t2_material_loading_memory_v1/QUALITY_REPORT.md`.
56
-
57
- ## Baselines
58
-
59
- Two compact PyTorch baselines were trained using complete-trajectory splits and
60
- train-only normalization. A pointwise MLP sees the current prescribed strain
61
- and material parameters but no loading history. A GRU sees the ordered strain
62
- history.
63
-
64
- | Baseline | Test RMSE | Test MAE | Test R2 |
65
  |---|---:|---:|---:|
66
- | Pointwise MLP | 244.54 MPa | 198.82 MPa | 0.4316 |
67
- | History-aware GRU | 16.52 MPa | 12.01 MPa | 0.9974 |
68
-
69
- The comparison demonstrates that the response is not a single-valued function
70
- of current strain and parameters; loading history contains essential predictive
71
- information. These are reproducible reference baselines, not claims of an
72
- optimal architecture.
73
-
74
- ![Baseline predictions](artifacts/t2_material_loading_memory_v1/baseline_predictions.png)
75
-
76
- ## Lightweight use
77
-
78
- The reader requires only NumPy and h5py:
79
-
80
- ```python
81
- from src.load_t2_material_loading_memory import load_trajectory
82
-
83
- sample = load_trajectory("data/t2_material_loading_memory_v1", 0)
84
- strain = sample["history"]["signed_equivalent_strain"]
85
- stress = sample["history"]["signed_equivalent_stress_pa"]
86
- print(sample["material_model"], sample["path_family"], strain.shape, stress.shape)
87
- ```
88
-
89
- ## Scope and limitations
90
-
91
- The data are synthetic three-dimensional small-strain material-point histories
92
- under prescribed proportional deviatoric strain. They are not structural
93
- finite-element fields, experimental material calibration or fatigue-life
94
- labels. Damage, temperature dependence, finite strain and non-proportional
95
- multiaxial loading are outside v1.
96
-
97
- Chaboche is explicitly marked as an experimental AgentFEM capability. Raw
98
- stress work on plastic strain is retained as a diagnostic, but it is not named
99
- thermodynamic dissipation because the current material-point contract does not
100
- expose a complete backstress storage/recovery energy split.
101
-
102
- ## Reproducibility and license
103
-
104
- The dataset was generated with AgentFEM `0.3.7.dev0`, exact commit
105
- `058faecc05aeda143d014fd229401003a9258bbb`. The frozen configuration,
106
- restartable generator, lightweight reader, baseline and tests are included.
107
-
108
- Dataset contents are released under CC BY 4.0. Included source code is released
109
- under Apache-2.0.
 
4
  - time-series-forecasting
5
  - other
6
  tags:
7
+ - mechanics
8
+ - finite-element-analysis
9
  - constitutive-modeling
10
+ - plasticity
11
+ - out-of-distribution
12
+ - physics-informed-machine-learning
 
13
  pretty_name: AgentFEM Material Loading Memory
14
  size_categories:
15
  - 1K<n<10K
16
  ---
17
 
18
+ # AgentFEM Material Loading Memory
19
 
20
+ An open, reproducible research dataset for path-dependent material modeling,
21
+ neural constitutive surrogates and finite-element deployment tests. The
22
+ repository now contains two compatible releases.
 
23
 
24
+ ## Releases
25
 
26
+ - **T2 v1:** 1,008 proportional cyclic J2 trajectories for loading-memory
27
+ baselines and forward/reversed ordering controls.
28
+ - **T2 multiaxial v2:** 1,024 independently generated J2 and Chaboche
29
+ trajectories in the complete five-dimensional deviatoric strain space,
30
+ with ID, path-OOD and parameter-OOD splits.
31
 
32
+ Start with `T2_V2_START_HERE.md` for the research extension or `START_HERE.md`
33
+ for the original v1 release.
 
 
 
 
 
 
 
 
34
 
35
+ ## Multiaxial v2 at a glance
 
 
 
36
 
37
+ - eight path families, including non-proportional, rotating and random
38
+ five-direction loading;
39
+ - 241 ordered states per trajectory;
40
+ - stress, strain, plastic strain, PEEQ, two-family total backstress, plastic
41
+ increments and physical diagnostics;
42
+ - zero quality-gate failures across all 1,024 trajectories;
43
+ - six model families evaluated under three frozen protocols;
44
+ - element-level deployment in a controlled notched-bar finite-element gate.
45
 
46
+ Stress RMSE on held-out trajectories:
 
 
 
 
 
 
47
 
48
+ | Model | ID | Path OOD | Parameter OOD |
 
 
 
 
 
 
 
 
 
 
49
  |---|---:|---:|---:|
50
+ | Pointwise MLP | 83.51 MPa | 105.81 MPa | 82.34 MPa |
51
+ | GRU | 17.26 MPa | 76.03 MPa | 19.26 MPa |
52
+ | LSTM | 22.67 MPa | 102.54 MPa | 22.89 MPa |
53
+ | Causal TCN | 21.82 MPa | 60.44 MPa | 21.99 MPa |
54
+ | Physics-state GRU | 23.06 MPa | 66.00 MPa | 24.19 MPa |
55
+ | Physics-integrator NN | **0.0128 MPa** | **0.0237 MPa** | **0.0275 MPa** |
56
+
57
+ ![Model comparison](artifacts/t2_multiaxial_ood_v2/model_protocol_comparison.png)
58
+
59
+ The physics-integrator NN predicts a bounded plastic-multiplier correction and
60
+ then applies differentiable J2/Chaboche consistency corrections while storing
61
+ plastic strain, PEEQ and two backstress tensors. It passed the mild and severe
62
+ structural gates with reaction relative-L2 errors of 1.65e-7 and 2.39e-7. The
63
+ severe case recorded one trust-region fallback at complete unloading.
64
+
65
+ This is a **white-box physics-fusion ceiling**: the embedded return-mapping
66
+ equations match the constitutive families that generated the synthetic data.
67
+ It demonstrates the benefit of architecture-level physics for this controlled
68
+ benchmark, but does not establish transfer to unknown materials or
69
+ misspecified physical models.
70
+
71
+ ## Reproduction and evidence
72
+
73
+ - Dataset details: `data/t2_multiaxial_ood_v2/README.md`
74
+ - Model card: `models/t2_multiaxial_ood_v2/README.md`
75
+ - Quality report: `artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md`
76
+ - Full metrics: `artifacts/t2_multiaxial_ood_v2/model_metrics.json`
77
+ - Structural gates: `artifacts/t2_multiaxial_ood_v2/structural_validation/`
78
+ - Manuscript roadmap: `docs/T2_MULTIAXIAL_RESEARCH_MEMO.md`
79
+ - Literature map: `docs/T2_MULTIAXIAL_LITERATURE.md`
80
+ - Commands: `T2_V2_REPRODUCE.md`
81
+
82
+ The data were generated with AgentFEM commit
83
+ `058faecc05aeda143d014fd229401003a9258bbb` (`0.3.7.dev0`). All quantities use
84
+ SI units. Voigt order is `xx, yy, zz, xy, yz, xz`, with tensor shear.
85
+
86
+ ## Scope
87
+
88
+ The releases are synthetic, small-strain material-point benchmarks. They do
89
+ not constitute experimental calibration, fatigue-life prediction, damage,
90
+ finite-strain plasticity or a production-certified learned material. Data are
91
+ CC BY 4.0; code files are covered by the included code license.
 
 
REPRODUCE.md ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproducing AgentFEM Material Loading Memory v1
2
+
3
+ ## Reproducibility levels
4
+
5
+ The release supports three distinct levels. Do not confuse them.
6
+
7
+ 1. **Data verification** checks the eight HDF5 shard hashes, schema, trajectory
8
+ IDs, splits, model labels, path-family counts and finite values. It requires
9
+ only Python, NumPy and h5py.
10
+ 2. **Baseline reproduction** retrains the published MLP and GRU from the frozen
11
+ trajectory splits. CPU training is sufficient.
12
+ 3. **Physics regeneration** recreates all trajectories with the exact AgentFEM
13
+ source revision, repackages the shards and reruns the numerical audit.
14
+
15
+ ## Obtain the release
16
+
17
+ Using the Hugging Face CLI:
18
+
19
+ ```bash
20
+ hf download HaomingLuo/AgentFEM-Material-Loading-Memory \
21
+ --repo-type dataset \
22
+ --local-dir AgentFEM-Material-Loading-Memory
23
+ cd AgentFEM-Material-Loading-Memory
24
+ ```
25
+
26
+ ## Verify the downloaded data
27
+
28
+ ```bash
29
+ conda env create -f environment-use.yml
30
+ conda activate agentfem-t2-use
31
+ bash reproduce_t2_v1.sh verify
32
+ ```
33
+
34
+ Expected summary:
35
+
36
+ - 1,008 unique trajectories;
37
+ - 768/120/120 train/validation/test trajectories;
38
+ - 504 J2 and 504 Chaboche trajectories;
39
+ - 168 trajectories in each of six path families;
40
+ - eight shard hashes matching `manifest.json`.
41
+
42
+ ## Reproduce the baselines
43
+
44
+ ```bash
45
+ bash reproduce_t2_v1.sh baseline
46
+ ```
47
+
48
+ The script trains with the published defaults: 40 MLP epochs and 60 GRU
49
+ epochs. Small numerical variation across PyTorch versions and hardware is
50
+ normal. Compare with `artifacts/t2_material_loading_memory_v1/baseline_metrics.json`,
51
+ not by requiring bitwise-identical neural-network weights.
52
+
53
+ ## Regenerate the physics data
54
+
55
+ The frozen physics environment uses Python 3.11, FEniCSx/DOLFINx 0.11.0 and
56
+ AgentFEM commit `058faecc05aeda143d014fd229401003a9258bbb`.
57
+
58
+ ```bash
59
+ conda env create -f environment-reproduce.yml
60
+ conda activate agentfem-t2-reproduce
61
+ bash reproduce_t2_v1.sh design
62
+ bash reproduce_t2_v1.sh full
63
+ ```
64
+
65
+ `design` checks the deterministic 1,008-case Sobol design without solving.
66
+ `full` performs the following steps:
67
+
68
+ 1. regenerate or resume all 1,008 material-point trajectories;
69
+ 2. build eight HDF5 shards;
70
+ 3. run analytical, constitutive and time-resolution audits;
71
+ 4. retrain both reference baselines;
72
+ 5. run the published automated tests.
73
+
74
+ The generator is restartable. For controlled parallel work, different workers
75
+ may run non-overlapping ranges before one worker packages and audits:
76
+
77
+ ```bash
78
+ python src/t2_material_loading_memory_v1.py --range 0 252
79
+ python src/t2_material_loading_memory_v1.py --range 252 504
80
+ python src/t2_material_loading_memory_v1.py --range 504 756
81
+ python src/t2_material_loading_memory_v1.py --range 756 1008
82
+ python src/t2_material_loading_memory_v1.py --package
83
+ python src/t2_material_loading_memory_v1.py --audit
84
+ ```
85
+
86
+ ## Exact source fallback
87
+
88
+ If the VCS entry in `environment-reproduce.yml` cannot install AgentFEM,
89
+ install it explicitly after creating the environment:
90
+
91
+ ```bash
92
+ git clone https://github.com/haoming-luo/agentfem.git
93
+ cd agentfem
94
+ git checkout 058faecc05aeda143d014fd229401003a9258bbb
95
+ python -m pip install .
96
+ cd ..
97
+ ```
98
+
99
+ Confirm before regeneration:
100
+
101
+ ```bash
102
+ python -c "import agentfem; print(agentfem.__version__)"
103
+ git -C agentfem rev-parse HEAD
104
+ ```
105
+
106
+ ## Reproducibility contract for extensions
107
+
108
+ - A sample is one complete trajectory; frames must not cross data splits.
109
+ - Keep SI units and record all material parameters and load-path controls.
110
+ - Preserve v1 as an immutable benchmark. Publish extensions under a new
111
+ dataset version and document migration rules.
112
+ - Failed or non-converged cases must be retained in a failure ledger rather
113
+ than silently removed.
114
+ - Split generation must be deterministic and grouped by trajectory and source
115
+ design.
116
+ - Report both predictive errors and physical-consistency errors.
SHA256SUMS CHANGED
@@ -1,14 +1,37 @@
1
  4a8e89fe65965e0b5ab34a6a1b8faa412666e426ce92500dba6af0c8f3e0ca0b CODE_LICENSE
2
  cdd59472c82b98d55f37288e0dd5994059194134fba6568029052eeff64faee4 DATA_LICENSE.md
3
- 0a423180979b86b932b153be5cd1538fb33a288643ab4a457b716916008b7933 README.md
 
 
 
 
 
 
4
  b06a19ffe36df25bbeedf35ccb8df9a9afff6df28870a354bf5c0dfc245fd948 artifacts/t2_material_loading_memory_v1/QUALITY_REPORT.md
5
  d19cc739e61edff4182ed700e924af7b051f24d37f07fbfa6581eff5e589483a artifacts/t2_material_loading_memory_v1/baseline_metrics.json
6
  0537f7cd76aae72abfa644761b8eec2f58903e66b8b2c040354f71650cfe0615 artifacts/t2_material_loading_memory_v1/baseline_models.pt
7
  9fc457b15d4e838a7dffd1d98d92fc002ee7d4e44f919c5ce1cf37e1bad75e8b artifacts/t2_material_loading_memory_v1/baseline_predictions.png
8
  d671538e1fe482de5ee12496185a7c2f65363d1376af4174016a902b1d8cb24d artifacts/t2_material_loading_memory_v1/hysteresis_preview.png
9
  4986bb10e002c1c2c1648b8ff3ee20676b2982215494e7f7eafaabdf7eeda298 artifacts/t2_material_loading_memory_v1/quality.json
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10
  5054d2c6e6d289c928e91c67b1c2535b7fc765d00ed7e05c39ec364c35981851 configs/t2_material_loading_memory_pilot.json
11
  50eb2dee418aba303e20ba6c219fb7d201d0a5a2391b933da2f3651e767191f5 configs/t2_material_loading_memory_v1.json
 
12
  d2523144f022f78b71c0f5b006c03a393123dc8bcad190e23375a994a248330d data/t2_material_loading_memory_v1/design.jsonl
13
  86ce6e7fa60f668736906c7158dfe0d5e3a8d40ca8a7e8d4e93df1d6a9afadfd data/t2_material_loading_memory_v1/index.jsonl
14
  0c863700bb310edbe6363ad3ca143015279226b7d1af9db25599ae79f7beb297 data/t2_material_loading_memory_v1/manifest.json
@@ -20,9 +43,78 @@ bf77420e04584151af10ea6e503fe4ab09946aa8c52d097ac8d4c902362387d3 data/t2_materi
20
  7039ab13425efac4e7bbdd415d31a5276acd3617a22163ad563c6817fca7360e data/t2_material_loading_memory_v1/shards/part-00005.h5
21
  67739a62d001131729ce506bb575a48555dd7dbbf393ac5a30cacf5094636854 data/t2_material_loading_memory_v1/shards/part-00006.h5
22
  37d8caa79a1cb672288f38e1625b4837671bfb9210416b10ce4a60ad5a912ef8 data/t2_material_loading_memory_v1/shards/part-00007.h5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
  d36bef7aa852e850a524e22f4b9689325b40320ad3046004d2f31df4c7605e26 src/baseline_t2_material_loading_memory.py
24
  21158e37d8742b20ef5817bb314fb0dfe926f9633b82087a063447d8e863f9fe src/load_t2_material_loading_memory.py
 
 
 
25
  79eddf3174eb343ef1e8dc4aef2df0a7be62819a03c594aaa22886fd49f57c2d src/t2_material_loading_memory.py
26
  604bc6283e7a461073d03a3b69440873d46809baf9e2ca4c79dfbeaa0704a5d7 src/t2_material_loading_memory_v1.py
 
 
 
 
 
27
  05c9cc11d32f54c88d3769b394b15f14bc283dc1f10b33b187b9e59166c676d6 tests/test_t2_material_loading_memory.py
28
  de16ddbac7c1e9dd020d2fff57da76d56ffd4ba2ce3b1ef61b41f02295c89e5a tests/test_t2_material_loading_memory_v1.py
 
 
 
 
 
1
  4a8e89fe65965e0b5ab34a6a1b8faa412666e426ce92500dba6af0c8f3e0ca0b CODE_LICENSE
2
  cdd59472c82b98d55f37288e0dd5994059194134fba6568029052eeff64faee4 DATA_LICENSE.md
3
+ 882fd0ac467a25c367798249f997bcc5460be8c5041b720f5e52809140ef8f60 KNOWN_LIMITATIONS.md
4
+ a0cbfe8420f8e7b9f27a5ec84c3b661a181112db356a8f0c4b4cb5dac66e7df5 PAPER_ROADMAP.md
5
+ 8b76aad385842a72927e3f561f879c12cfbfa369d7f209cb1947214676ddd325 README.md
6
+ 5884a4ae4cb5fd092b6dec45bd61107cf2365fa0500e4c1ff34a051583a16148 REPRODUCE.md
7
+ e552710f9fea8e21877e2011a457f75c790968fd5dbfd1549f066aa567875ea4 START_HERE.md
8
+ 9e62aabbc8bd24dd38539b03bf5f9c1e73507bb9a9d1b73c37c0d3ca6b2f7f14 T2_V2_REPRODUCE.md
9
+ c1329e03b40e57f66370343bf3f8a078149fefc28fbfbab80ee00412124571f7 T2_V2_START_HERE.md
10
  b06a19ffe36df25bbeedf35ccb8df9a9afff6df28870a354bf5c0dfc245fd948 artifacts/t2_material_loading_memory_v1/QUALITY_REPORT.md
11
  d19cc739e61edff4182ed700e924af7b051f24d37f07fbfa6581eff5e589483a artifacts/t2_material_loading_memory_v1/baseline_metrics.json
12
  0537f7cd76aae72abfa644761b8eec2f58903e66b8b2c040354f71650cfe0615 artifacts/t2_material_loading_memory_v1/baseline_models.pt
13
  9fc457b15d4e838a7dffd1d98d92fc002ee7d4e44f919c5ce1cf37e1bad75e8b artifacts/t2_material_loading_memory_v1/baseline_predictions.png
14
  d671538e1fe482de5ee12496185a7c2f65363d1376af4174016a902b1d8cb24d artifacts/t2_material_loading_memory_v1/hysteresis_preview.png
15
  4986bb10e002c1c2c1648b8ff3ee20676b2982215494e7f7eafaabdf7eeda298 artifacts/t2_material_loading_memory_v1/quality.json
16
+ 722c22826844dd451b28d938f72c09f0ee17a13cb7f599a8c9691570167273ad artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md
17
+ 54b95a1e83f0836a460370420e0eaed8653818f56252b5e2745739c42137be42 artifacts/t2_multiaxial_ood_v2/model_metrics.json
18
+ 934b3cb6c5092849141f18ae7914ad58463b4f99ce4accd3a8d8eb390fca9a95 artifacts/t2_multiaxial_ood_v2/model_protocol_comparison.png
19
+ 4651c319a4d4bfab88efb42b54467619b03c9a14175e1eeffb94ad7216b25fa4 artifacts/t2_multiaxial_ood_v2/multiaxial_hysteresis_preview.png
20
+ 7aae708be60dbb7a41b3c16068fd4eebc8cf4c5cabfe8f09ec417843f252bb18 artifacts/t2_multiaxial_ood_v2/physics_integrator_summary.json
21
+ f70ebec439c8972772649e621c18def3f94c9fe6b832d4e8e9cf8801ebffa0c0 artifacts/t2_multiaxial_ood_v2/quality.json
22
+ 69b803b9a1449bda3e8208fe27bcc1e868e4408b8582c2fb9c56a128922e971c artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/notched_bar_comparison.png
23
+ 8eb1650c1358f618f02bd51e668c23dfe57f12555a9f8c8fa97df25f3c9cd889 artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/notched_bar_results.npz
24
+ 3f13f49c24a0fd69e82710001a99aed812faceca330c11dedce6e2e48d511aaa artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/structural_validation.json
25
+ d7af84c60b140016d5abba35454a63d84bd58584911d0496760d35b9f276c01f artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/notched_bar_comparison.png
26
+ ded38f99016820dacb06d1a325b7cf6fd13a1d7053020803ff6c45ec2c2ec268 artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/notched_bar_results.npz
27
+ 9b2d02481d691954fa2622cbf9608a3bc68858ccec44732b69b5bfea0c347f5e artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/structural_validation.json
28
+ 740e645838d4b3276169a72e304aee7cab054dab8a63edb60d5867a87d63af57 artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/notched_bar_comparison.png
29
+ 999ef1d758cedec3d918662f223448826ff037b86423821f99d7f3e8bb959d51 artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/notched_bar_results.npz
30
+ 21491c8fe4848c20149f2f25b4baad4bda35f5806fd002606bbbf2b0798a4c3f artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/structural_validation.json
31
+ 2929ddd73ad5e2bb1258aa5501c89a41c2c1ce3548c5a9a84aa133e12b4ab1c2 artifacts/t2_multiaxial_ood_v2/training_summary.json
32
  5054d2c6e6d289c928e91c67b1c2535b7fc765d00ed7e05c39ec364c35981851 configs/t2_material_loading_memory_pilot.json
33
  50eb2dee418aba303e20ba6c219fb7d201d0a5a2391b933da2f3651e767191f5 configs/t2_material_loading_memory_v1.json
34
+ 622237a1fe57b695ccdfc5b9d2a28d82d4e6be051be5443cd89537a0a091cbf6 configs/t2_multiaxial_ood_v2.json
35
  d2523144f022f78b71c0f5b006c03a393123dc8bcad190e23375a994a248330d data/t2_material_loading_memory_v1/design.jsonl
36
  86ce6e7fa60f668736906c7158dfe0d5e3a8d40ca8a7e8d4e93df1d6a9afadfd data/t2_material_loading_memory_v1/index.jsonl
37
  0c863700bb310edbe6363ad3ca143015279226b7d1af9db25599ae79f7beb297 data/t2_material_loading_memory_v1/manifest.json
 
43
  7039ab13425efac4e7bbdd415d31a5276acd3617a22163ad563c6817fca7360e data/t2_material_loading_memory_v1/shards/part-00005.h5
44
  67739a62d001131729ce506bb575a48555dd7dbbf393ac5a30cacf5094636854 data/t2_material_loading_memory_v1/shards/part-00006.h5
45
  37d8caa79a1cb672288f38e1625b4837671bfb9210416b10ce4a60ad5a912ef8 data/t2_material_loading_memory_v1/shards/part-00007.h5
46
+ c0915b6d8bed2c3bd2d5ae156ff42e6c491b81eb6fc797614c6124f7485b1184 data/t2_multiaxial_ood_v2/README.md
47
+ 6eee820f46b173f927de5b5033a18b2e9e76c03f019c43f8fc505248b875a312 data/t2_multiaxial_ood_v2/design.jsonl
48
+ a13c1a19098d42d8729184b975e96bbf911c340367c5b46d9aa87887b22268ba data/t2_multiaxial_ood_v2/index.jsonl
49
+ 99793bdd23bb2abebc8d84ab2e2b7875d6c48c1ca62287d4cd83bff38dec83b6 data/t2_multiaxial_ood_v2/manifest.json
50
+ 16456ca1050a8afff4b687610d0af7b952dcd2b82149ad61a13e7f1296daa1d4 data/t2_multiaxial_ood_v2/shards/part-00000.h5
51
+ 3d56b126367a7cc2e7c7adeca5ae7badaafc91ad4130d5cab444a18c81ec3b10 data/t2_multiaxial_ood_v2/shards/part-00001.h5
52
+ 7742b2533ae3a7c2f5018e83abfa44c236caed8c11766d357f56dc92b136f84c data/t2_multiaxial_ood_v2/shards/part-00002.h5
53
+ 189d2ef804c3639fb85a58bddacc982e2d7634add25f0ea214c2652f69193aa6 data/t2_multiaxial_ood_v2/shards/part-00003.h5
54
+ 364cdd35f7d8280992267ceda7c4313a197811f8df408763c1f60a144a2a0c59 data/t2_multiaxial_ood_v2/shards/part-00004.h5
55
+ bf3abef7c1a1c1686725014f5be2b49189bca023014d889d418c8950a7699e80 data/t2_multiaxial_ood_v2/shards/part-00005.h5
56
+ 27fd47d03fff7ca8e394be61d4a4af2366de9d38ed695e47cbb2ae1fa8c90af7 data/t2_multiaxial_ood_v2/shards/part-00006.h5
57
+ 9d1b9d2bc4770cd67516b2f56ba56ae951d6d072a0f0286efccd18c6e94a686e data/t2_multiaxial_ood_v2/shards/part-00007.h5
58
+ 961e1520be9d9f6277423f5f2d422ccc9fde1a7c680d65a66724107edad43365 docs/T2_MULTIAXIAL_LITERATURE.md
59
+ 46dd7c86bf01b1faffb6216b171452ead2fc5c449512e34683f21ed6c907fd01 docs/T2_MULTIAXIAL_RESEARCH_MEMO.md
60
+ bcb156b8ecc18d5f222e9671dcf792a142fa7969b0eb9455591506a66ed21c36 environment-reproduce.yml
61
+ f841ccfb11c917545824bb4aaae605f39ce16aa1441f2a6ca182b5c8951c193a environment-use.yml
62
+ 2b0b7be156eae06668a54ab20ad40a7b16dc0d3ce27083096003f188a64054bd models/t2_multiaxial_ood_v2/README.md
63
+ c078b5514c8fc2245be0dcb5a7051a9e229f4168dc4ace5753ef8bbc6dd775d4 models/t2_multiaxial_ood_v2/id_causal_tcn.pt
64
+ a28b9731ef309c31d721086e5c4d44a4fcedebb3f2284d42dbfc4ca99bee1678 models/t2_multiaxial_ood_v2/id_causal_tcn_history.json
65
+ b346b3f373d5bea97febc396cb0f4f4c8db59041c97dcea345e975624605d0fb models/t2_multiaxial_ood_v2/id_gru.pt
66
+ 5dbe2d240055d660b1f2b65d6b795c14d6164674a907af066454d8730dc5927a models/t2_multiaxial_ood_v2/id_gru_history.json
67
+ c058e02c5283f9367b97c68bcdcf66d9319462a2d60b23e7587162a09cd1da5b models/t2_multiaxial_ood_v2/id_lstm.pt
68
+ f297cf91466d3fe8f8efb48ff38f55b74c3f95e80a3d32cc5b38dcd5e6ba373d models/t2_multiaxial_ood_v2/id_lstm_history.json
69
+ 4c682f1bc1ec6c8d2db5f542367fbd6daa4922ef566387d2f35974fbd669027b models/t2_multiaxial_ood_v2/id_normalization.json
70
+ e46c537a2963fbe6bea4c215f4118ca0b496ebfda8f5e2c0cf31a5ba455b0430 models/t2_multiaxial_ood_v2/id_physics_integrator_nn.pt
71
+ b535a1dd4e4682a96a9d5d5b9e3405ef16b058956763f66125bc3f03d76010f6 models/t2_multiaxial_ood_v2/id_physics_integrator_nn_history.json
72
+ 600e6e021e63990c005cab724397d42eaf49b3c0f06a3e6974bf44c6fe542498 models/t2_multiaxial_ood_v2/id_physics_state_gru.pt
73
+ 56e443773f51c475d624041e169eb170a6552e357e905f56cfd84d727d881ebd models/t2_multiaxial_ood_v2/id_physics_state_gru_history.json
74
+ e46bcadac4daaee66b2efdba5f790e9e28a330036172aa65ae1f4cc14df4d47a models/t2_multiaxial_ood_v2/id_pointwise_mlp.pt
75
+ 0e14ef379256b97176cf73fd4c83754d8ef956de4ef51f3b3d096478096a8f7c models/t2_multiaxial_ood_v2/id_pointwise_mlp_history.json
76
+ ced5a5222c7656c1347b8f8c7380f050ef9d2f758e8548e0ce28b69bdc65c6b5 models/t2_multiaxial_ood_v2/parameter_ood_causal_tcn.pt
77
+ 55b6f565a9a5b0b61bbba3a7dcc0572c8572f19939b07a036e791b4e60a76153 models/t2_multiaxial_ood_v2/parameter_ood_causal_tcn_history.json
78
+ da578b8b562530dbe6425caaa9cdbab2230957161258c11a7a2b9a204d1ab029 models/t2_multiaxial_ood_v2/parameter_ood_gru.pt
79
+ f5c785caf4fa75943b3451aeb051a375ca5a1b00abaed603a3e4ee15c50ef254 models/t2_multiaxial_ood_v2/parameter_ood_gru_history.json
80
+ c267359a397e0c27ee11a4f220c2ec8eef8390e70389e92bcb0f4d1507dbed52 models/t2_multiaxial_ood_v2/parameter_ood_lstm.pt
81
+ 700b2a2035da1090f74a4355678c7970f63b98582674cc720cb80bb130b2ac56 models/t2_multiaxial_ood_v2/parameter_ood_lstm_history.json
82
+ 1526e6ab0f9c5a9641bf50443445cb072ace38536610bc7c1762bb751e71873e models/t2_multiaxial_ood_v2/parameter_ood_normalization.json
83
+ cf8f6a75d1356e08ccf2439f93f49c4dc629456974eeba1159572f198755313c models/t2_multiaxial_ood_v2/parameter_ood_physics_integrator_nn.pt
84
+ 6137632878dd2ec15cf356ebe0dd486bb708453c8a241cee96d782579bc8e612 models/t2_multiaxial_ood_v2/parameter_ood_physics_integrator_nn_history.json
85
+ 59db7abe20857c7dcbe157b9d702bc71958771692fe0ea9f391dde8efeca74ea models/t2_multiaxial_ood_v2/parameter_ood_physics_state_gru.pt
86
+ 7fb343b9efdbaca64e9e06ecd52df7900c248c99c9782aed393537d49bbbee86 models/t2_multiaxial_ood_v2/parameter_ood_physics_state_gru_history.json
87
+ bfd9dd42ffb36a38c6591ed5c68eaef47e19c580b83669ca68c5ac8bc1eaaf30 models/t2_multiaxial_ood_v2/parameter_ood_pointwise_mlp.pt
88
+ 230bedd7aab23020f42f956b3ba550fbb43e0da8a4adfad04d9baa954a3b91bd models/t2_multiaxial_ood_v2/parameter_ood_pointwise_mlp_history.json
89
+ 4f57c8c7cd668fec6224689574a744d5ed27f9ce3d6ba89388a2b869d12d85b4 models/t2_multiaxial_ood_v2/path_ood_causal_tcn.pt
90
+ 522fc250a90575319caef23831bbef3a26fa5b2d3fc6f6c457a7c13bb1e44cf0 models/t2_multiaxial_ood_v2/path_ood_causal_tcn_history.json
91
+ db6dd4f4f314cefc83a0cb2317bd22dc571ad49ce8945d968bcf6a9d4039618a models/t2_multiaxial_ood_v2/path_ood_gru.pt
92
+ d41faf4dab2173f672dae3d899ce00f15c372f4ee32e2aa0303a496fe3b451a1 models/t2_multiaxial_ood_v2/path_ood_gru_history.json
93
+ ee2a768fa9e4cbb291e2bd39f20366da297a3f18765d7f3b5b87c93dae5bb51b models/t2_multiaxial_ood_v2/path_ood_lstm.pt
94
+ 255dfbf87c1a6cd1d87dba06d4ecdba36522a99831928ae9ac3d8ef91d9cf904 models/t2_multiaxial_ood_v2/path_ood_lstm_history.json
95
+ 99754f29cc5cc085716ccc1549c897d833589d17bd343c4da732f375257a7082 models/t2_multiaxial_ood_v2/path_ood_normalization.json
96
+ 8177fdd618367c478b1108ec0fad90be44a18df9ff1f9af3b719722ffa3a8e28 models/t2_multiaxial_ood_v2/path_ood_physics_integrator_nn.pt
97
+ 7ba19d106d92a0c6308c3cfae0bb0ebc1cbeea0924e1c967ac82e8a85f768de3 models/t2_multiaxial_ood_v2/path_ood_physics_integrator_nn_history.json
98
+ 14cb4597811838917c98e148fd1570881d4383771b897257e7bacc6ca0e724c8 models/t2_multiaxial_ood_v2/path_ood_physics_state_gru.pt
99
+ d88b52443a7c0074516eddf5860c40cd40b488d872b905bac49c08f7980e2869 models/t2_multiaxial_ood_v2/path_ood_physics_state_gru_history.json
100
+ e4bb90d495371ab8bc8bb5883cc5bcc6ececd3f62372b221ca6b4e98558d86fe models/t2_multiaxial_ood_v2/path_ood_pointwise_mlp.pt
101
+ 5356d05b393fe7f8a09bd00c28466c965f7c2260b7b6ae05a11c59d75b0b54c9 models/t2_multiaxial_ood_v2/path_ood_pointwise_mlp_history.json
102
+ 9098c128e0b540ab83b1f1323d1c3b437a8130a8a88309684fd931c09ca6d4fe reproduce_t2_v1.sh
103
  d36bef7aa852e850a524e22f4b9689325b40320ad3046004d2f31df4c7605e26 src/baseline_t2_material_loading_memory.py
104
  21158e37d8742b20ef5817bb314fb0dfe926f9633b82087a063447d8e863f9fe src/load_t2_material_loading_memory.py
105
+ 229915f1d28cceb7e63eb9795b0e9e640eae8c4ff38052c3374a5ef7d89ffd3b src/package_t2_material_loading_memory_v1.py
106
+ 3521afa9ce4ebb5a9c2cdcf869612989057f9ba7e4b44101facda95d2764e09a src/package_t2_multiaxial_release.py
107
+ f6d80c48d66c2d5e14fdccff65d9fecd8a632f5322f68eb5f9ec23679b8f67ed src/structural_validate_t2_models.py
108
  79eddf3174eb343ef1e8dc4aef2df0a7be62819a03c594aaa22886fd49f57c2d src/t2_material_loading_memory.py
109
  604bc6283e7a461073d03a3b69440873d46809baf9e2ca4c79dfbeaa0704a5d7 src/t2_material_loading_memory_v1.py
110
+ 26da999c809d605c3ca710af1d765bff296ec8e03623eb0feff62deaf5a94d86 src/t2_multiaxial_ood_v2.py
111
+ 27653364825e5f6d64edb91c7669c40257ecfa839bdcc987f6150b7376706b9a src/train_t2_multiaxial_models.py
112
+ 50b761b008c07ae8419db652a37270ebe15d5c9b7168d87cef1f836c9acc245f src/train_t2_physics_integrator.py
113
+ 2efb9502d85dde486939d02613e13fe6d254e8c6e74e2d82f2eb3201c01d1ac7 src/verify_t2_download.py
114
+ 79319ee3dc4c32abc5b9cf68ce5efd226c82d36c43cd74a26a7f49dddffe9154 src/verify_t2_multiaxial_download.py
115
  05c9cc11d32f54c88d3769b394b15f14bc283dc1f10b33b187b9e59166c676d6 tests/test_t2_material_loading_memory.py
116
  de16ddbac7c1e9dd020d2fff57da76d56ffd4ba2ce3b1ef61b41f02295c89e5a tests/test_t2_material_loading_memory_v1.py
117
+ f851886f02ce4e30c24881ef167ebb0462616a18d486ccf4d384afe61146dab2 tests/test_t2_multiaxial_models.py
118
+ f135a64d401ca69f2774fca56d187b3570b733e339f272f13b6a06d3fc0d1d20 tests/test_t2_multiaxial_ood_v2.py
119
+ b2753c68fb0b1819f281828cc84e741566453e70019243acedf51dd07ed997d3 tests/test_t2_physics_integrator.py
120
+ 5e391f7cdcdf4ba800bb87d6d1c2db58131303f91376df1aad78bcc2eb47149c tests/test_t2_structural_validation.py
START_HERE.md ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Start here: AgentFEM Material Loading Memory v1
2
+
3
+ This repository is a research handoff, not only a file archive. It contains the
4
+ published dataset, the exact generation configuration, reusable readers,
5
+ reference baselines, numerical tests, quality evidence and a concrete route for
6
+ extending the work into a research article.
7
+
8
+ ## Choose your route
9
+
10
+ ### 1. Use the data or train a new model
11
+
12
+ You do **not** need AgentFEM. Create the lightweight environment and verify the
13
+ download:
14
+
15
+ ```bash
16
+ conda env create -f environment-use.yml
17
+ conda activate agentfem-t2-use
18
+ bash reproduce_t2_v1.sh verify
19
+ ```
20
+
21
+ Start with `src/load_t2_material_loading_memory.py`. The current HDF5 data can
22
+ be used directly for GRU, LSTM, transformer, state-space or operator-learning
23
+ experiments.
24
+
25
+ ### 2. Reproduce the published baseline
26
+
27
+ ```bash
28
+ bash reproduce_t2_v1.sh baseline
29
+ ```
30
+
31
+ This verifies the data and retrains the pointwise MLP and history-aware GRU.
32
+ The published reference metrics are in
33
+ `artifacts/t2_material_loading_memory_v1/baseline_metrics.json`.
34
+
35
+ ### 3. Regenerate the physics data
36
+
37
+ Use `environment-reproduce.yml`, which installs FEniCSx and the exact AgentFEM
38
+ commit used for v1:
39
+
40
+ ```bash
41
+ conda env create -f environment-reproduce.yml
42
+ conda activate agentfem-t2-reproduce
43
+ bash reproduce_t2_v1.sh design
44
+ bash reproduce_t2_v1.sh full
45
+ ```
46
+
47
+ Full regeneration is deterministic but rewrites generated shards and baseline
48
+ artifacts in the working copy. See `REPRODUCE.md` before running it.
49
+
50
+ ## What is already established
51
+
52
+ - 1,008 complete trajectories, not 121,968 independent frames.
53
+ - Two explicitly labelled constitutive models: J2 linear isotropic hardening
54
+ and Chaboche combined hardening.
55
+ - Six balanced proportional loading-path families.
56
+ - Complete-trajectory train/validation/test splits with no frame leakage.
57
+ - Numerical consistency, analytical, refinement and checksum evidence.
58
+ - A strong history-aware baseline and a deliberately history-blind control.
59
+
60
+ ## What remains for a strong paper
61
+
62
+ The current release is a verified foundation, not the full paper claim. The
63
+ recommended first paper studies generalization of learned path-dependent
64
+ constitutive models. Its priority additions are true non-proportional
65
+ multiaxial paths, physically meaningful out-of-distribution splits,
66
+ physics-aware metrics, stronger sequence-model comparisons and deployment in a
67
+ small structural finite-element example. See `PAPER_ROADMAP.md`.
68
+
69
+ ## Source of truth and citation
70
+
71
+ - Dataset: <https://huggingface.co/datasets/HaomingLuo/AgentFEM-Material-Loading-Memory>
72
+ - AgentFEM: <https://github.com/haoming-luo/agentfem>
73
+ - AgentFEM commit used for v1:
74
+ `058faecc05aeda143d014fd229401003a9258bbb`
75
+ - Dataset license: CC BY 4.0
76
+ - Included source-code license: Apache-2.0
77
+
78
+ When extending the dataset, preserve trajectory-level grouping, units, model
79
+ labels, random seeds, quality failures and the original v1 test set. Record a
80
+ new schema or dataset version whenever fields or physical scope change.
T2_V2_REPRODUCE.md ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproduce T2 multiaxial OOD v2
2
+
3
+ ## Verify downloaded data
4
+
5
+ ```bash
6
+ python src/verify_t2_multiaxial_download.py data/t2_multiaxial_ood_v2
7
+ ```
8
+
9
+ ## Regenerate the data
10
+
11
+ Use Python 3.11, DOLFINx 0.11.0 and AgentFEM commit
12
+ `058faecc05aeda143d014fd229401003a9258bbb`.
13
+
14
+ ```bash
15
+ python src/t2_multiaxial_ood_v2.py --validate-design
16
+ python src/t2_multiaxial_ood_v2.py --all
17
+ ```
18
+
19
+ The restartable generator also accepts non-overlapping `--range START STOP`
20
+ jobs, followed by `--package` and `--audit`.
21
+
22
+ ## Retrain all models
23
+
24
+ ```bash
25
+ python src/train_t2_multiaxial_models.py --epochs 35 --batch-size 32
26
+ python src/train_t2_multiaxial_models.py \
27
+ --epochs 45 --batch-size 32 --models physics_state_gru --seed 20261011
28
+ python src/train_t2_physics_integrator.py --epochs 40
29
+ ```
30
+
31
+ The second command replaces the preliminary soft-constraint physics models
32
+ with the released hard-monotonic-state checkpoints.
33
+
34
+ ## Reproduce structural gates
35
+
36
+ ```bash
37
+ python src/structural_validate_t2_models.py --case mild_id
38
+ python src/structural_validate_t2_models.py --case severe_ood
39
+ ```
40
+
41
+ The severe physics-state GRU is expected to record a failed global solve at
42
+ load step 58. The physics-integrator NN passes after recording one global
43
+ trust-region fallback at the same complete-unloading step; do not hide or
44
+ remove this fallback from comparisons.
45
+
46
+ ## Run tests
47
+
48
+ ```bash
49
+ python -m pytest -q \
50
+ tests/test_t2_multiaxial_ood_v2.py \
51
+ tests/test_t2_multiaxial_models.py \
52
+ tests/test_t2_physics_integrator.py \
53
+ tests/test_t2_structural_validation.py
54
+ ```
T2_V2_START_HERE.md ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Start here: T2 multiaxial OOD research release
2
+
3
+ This package is designed for a collaborator or AI agent to continue the
4
+ research without reconstructing the project history.
5
+
6
+ ## Read in this order
7
+
8
+ 1. `README.md` for the complete dataset overview.
9
+ 2. `artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md` for physics validation.
10
+ 3. `artifacts/t2_multiaxial_ood_v2/model_metrics.json` for all model results.
11
+ 4. `models/t2_multiaxial_ood_v2/README.md` before loading checkpoints.
12
+ 5. `docs/T2_MULTIAXIAL_RESEARCH_MEMO.md` for the manuscript argument and next
13
+ method contribution.
14
+ 6. `T2_V2_REPRODUCE.md` for exact commands.
15
+
16
+ ## What may be claimed now
17
+
18
+ - The data generation and physical audits passed.
19
+ - History-aware models strongly outperform a history-blind control on ID data.
20
+ - Path-family OOD is substantially harder than the frozen parameter-OOD test.
21
+ - Selected hard physical constraints remove corresponding state violations.
22
+ - A state-output physics GRU did not transfer reliably to the structural gate.
23
+ - A physics-embedded neural stress integrator reduced path-OOD RMSE to 0.0237
24
+ MPa and passed both frozen structural gates; the severe case recorded one
25
+ trust-region fallback at complete unloading.
26
+
27
+ ## What may not be claimed now
28
+
29
+ - production-ready neural constitutive modeling;
30
+ - general 3D AgentFEM learned-material integration;
31
+ - experimental-material accuracy;
32
+ - transfer of the white-box result to unknown or misspecified material laws;
33
+ - general structural convergence without the recorded global fallback;
34
+ - end-to-end acceleration in the current Python structural demonstrator.
35
+
36
+ ## Highest-value continuation
37
+
38
+ Develop a gray-box evolution model in which the embedded constitutive skeleton
39
+ is deliberately incomplete and the network learns missing evolution terms or
40
+ latent internal variables. Add a verified consistent tangent, 2D/3D gates and
41
+ experimental or high-fidelity microstructural data. Retain the current splits
42
+ and structural gates unchanged so improvement is measurable.
artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # T2 multiaxial OOD v2 quality report
2
+
3
+ Status: **accepted**
4
+ Independent trajectories: 1024
5
+ Quality failures: 0
6
+
7
+ ## Physics and numerical evidence
8
+
9
+ - Maximum yield-surface relative residual: 1.575e-11
10
+ - Maximum associated-flow alignment error: 6.661e-16
11
+ - Maximum rigid-rotation covariance error: 3.610e-15
12
+ - Maximum 241-to-481-state stress change: 1.466%
13
+ - Maximum 241-to-481-state PEEQ change: 0.446%
14
+ - Minimum non-proportional path rank ratio: 4.241e-01
15
+
16
+ The path, parameter and ID protocols are independent labels over the same immutable trajectories.
17
+ They must be trained and reported separately. This synthetic release is not experimental calibration,
18
+ fatigue-life data, damage evolution or a structural finite-element field dataset.
19
+
20
+ ![Multiaxial preview](multiaxial_hysteresis_preview.png)
artifacts/t2_multiaxial_ood_v2/model_metrics.json ADDED
@@ -0,0 +1,290 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "id": {
3
+ "causal_tcn": {
4
+ "best_validation_loss": 0.052073314785957336,
5
+ "decimated_121_state_rmse_mpa": 50.41107940673828,
6
+ "epochs_completed": 35,
7
+ "inference_seconds": 0.10764549998566508,
8
+ "mae_mpa": 13.459792137145996,
9
+ "mean_peak_mises_relative_error": 0.09287908673286438,
10
+ "mean_work_relative_error": 0.06005401164293289,
11
+ "mean_yield_surface_relative_residual_using_reference_state": 0.11287368834018707,
12
+ "parameter_count": 64878,
13
+ "r2": 0.9642560482025146,
14
+ "rmse_mpa": 21.822433471679688,
15
+ "trajectory_count": 128
16
+ },
17
+ "gru": {
18
+ "best_validation_loss": 0.03528529033064842,
19
+ "decimated_121_state_rmse_mpa": 27.9344482421875,
20
+ "epochs_completed": 35,
21
+ "inference_seconds": 0.06031920784153044,
22
+ "mae_mpa": 10.743799209594727,
23
+ "mean_peak_mises_relative_error": 0.06271586567163467,
24
+ "mean_work_relative_error": 0.05553208291530609,
25
+ "mean_yield_surface_relative_residual_using_reference_state": 0.08399201184511185,
26
+ "parameter_count": 51846,
27
+ "r2": 0.9776403903961182,
28
+ "rmse_mpa": 17.25974464416504,
29
+ "trajectory_count": 128
30
+ },
31
+ "lstm": {
32
+ "best_validation_loss": 0.0651465654373169,
33
+ "decimated_121_state_rmse_mpa": 41.104408264160156,
34
+ "epochs_completed": 35,
35
+ "inference_seconds": 0.08951037493534386,
36
+ "mae_mpa": 13.053536415100098,
37
+ "mean_peak_mises_relative_error": 0.06156584620475769,
38
+ "mean_work_relative_error": 0.06417670100927353,
39
+ "mean_yield_surface_relative_residual_using_reference_state": 0.09419497102499008,
40
+ "parameter_count": 68982,
41
+ "r2": 0.9614183902740479,
42
+ "rmse_mpa": 22.6721248626709,
43
+ "trajectory_count": 128
44
+ },
45
+ "physics_integrator_nn": {
46
+ "backstress_rmse_mpa": 1.1107648788311053e-05,
47
+ "best_validation_loss": 0.0002665960637386888,
48
+ "decimated_121_state_rmse_mpa": 1.0767621994018555,
49
+ "epochs_completed": 40,
50
+ "inference_seconds": 0.1447634168434888,
51
+ "mae_mpa": 0.004262915346771479,
52
+ "maximum_plastic_strain_trace": 6.51925802230835e-09,
53
+ "mean_consistency_relative_residual": 1.493731183188629e-08,
54
+ "mean_peak_mises_relative_error": 3.548976019374095e-05,
55
+ "mean_work_relative_error": 4.189212995697744e-05,
56
+ "parameter_count": 3265,
57
+ "peeq_nonmonotone_fraction": 0.0,
58
+ "peeq_rmse": 7.873447884776397e-07,
59
+ "plastic_strain_rmse": 8.518962602011015e-08,
60
+ "r2": 1.0,
61
+ "rmse_mpa": 0.012793608009815216,
62
+ "trajectory_count": 128
63
+ },
64
+ "physics_state_gru": {
65
+ "backstress_rmse_mpa": 24.632232666015625,
66
+ "best_validation_loss": 0.0608750656247139,
67
+ "decimated_121_state_rmse_mpa": 34.93246841430664,
68
+ "epochs_completed": 45,
69
+ "inference_seconds": 0.09351545805111527,
70
+ "mae_mpa": 14.640517234802246,
71
+ "maximum_plastic_strain_trace": 0.0,
72
+ "mean_peak_mises_relative_error": 0.18708987534046173,
73
+ "mean_work_relative_error": 0.047915033996105194,
74
+ "mean_yield_surface_relative_residual_using_reference_state": 0.13221898674964905,
75
+ "parameter_count": 63771,
76
+ "peeq_nonmonotone_fraction": 0.0,
77
+ "peeq_rmse": 0.0042837150394916534,
78
+ "plastic_strain_rmse": 0.00016164015687536448,
79
+ "r2": 0.9600764513015747,
80
+ "rmse_mpa": 23.063045501708984,
81
+ "trajectory_count": 128
82
+ },
83
+ "pointwise_mlp": {
84
+ "best_validation_loss": 0.6067270636558533,
85
+ "decimated_121_state_rmse_mpa": 83.54443359375,
86
+ "epochs_completed": 35,
87
+ "inference_seconds": 0.004329124931246042,
88
+ "mae_mpa": 52.44876480102539,
89
+ "mean_peak_mises_relative_error": 0.2378098964691162,
90
+ "mean_work_relative_error": 0.8128710985183716,
91
+ "mean_yield_surface_relative_residual_using_reference_state": 0.5040972232818604,
92
+ "parameter_count": 11718,
93
+ "r2": 0.4765147566795349,
94
+ "rmse_mpa": 83.51302337646484,
95
+ "trajectory_count": 128
96
+ }
97
+ },
98
+ "parameter_ood": {
99
+ "causal_tcn": {
100
+ "best_validation_loss": 0.06454462558031082,
101
+ "decimated_121_state_rmse_mpa": 50.147891998291016,
102
+ "epochs_completed": 35,
103
+ "inference_seconds": 0.08833870780654252,
104
+ "mae_mpa": 13.355766296386719,
105
+ "mean_peak_mises_relative_error": 0.09955132752656937,
106
+ "mean_work_relative_error": 0.06557949632406235,
107
+ "mean_yield_surface_relative_residual_using_reference_state": 0.12519554793834686,
108
+ "parameter_count": 64878,
109
+ "r2": 0.9634166955947876,
110
+ "rmse_mpa": 21.985719680786133,
111
+ "trajectory_count": 104
112
+ },
113
+ "gru": {
114
+ "best_validation_loss": 0.03666621819138527,
115
+ "decimated_121_state_rmse_mpa": 28.72995948791504,
116
+ "epochs_completed": 35,
117
+ "inference_seconds": 0.05565433297306299,
118
+ "mae_mpa": 11.537033081054688,
119
+ "mean_peak_mises_relative_error": 0.06023126468062401,
120
+ "mean_work_relative_error": 0.06329581141471863,
121
+ "mean_yield_surface_relative_residual_using_reference_state": 0.09963684529066086,
122
+ "parameter_count": 51846,
123
+ "r2": 0.9719248414039612,
124
+ "rmse_mpa": 19.260169982910156,
125
+ "trajectory_count": 104
126
+ },
127
+ "lstm": {
128
+ "best_validation_loss": 0.056471772491931915,
129
+ "decimated_121_state_rmse_mpa": 41.76096725463867,
130
+ "epochs_completed": 35,
131
+ "inference_seconds": 0.08064724993892014,
132
+ "mae_mpa": 13.136635780334473,
133
+ "mean_peak_mises_relative_error": 0.06535962224006653,
134
+ "mean_work_relative_error": 0.0623355470597744,
135
+ "mean_yield_surface_relative_residual_using_reference_state": 0.10602641105651855,
136
+ "parameter_count": 68982,
137
+ "r2": 0.9603530168533325,
138
+ "rmse_mpa": 22.887819290161133,
139
+ "trajectory_count": 104
140
+ },
141
+ "physics_integrator_nn": {
142
+ "backstress_rmse_mpa": 1.0817180736921728e-05,
143
+ "best_validation_loss": 0.00028723079594783485,
144
+ "decimated_121_state_rmse_mpa": 1.1382688283920288,
145
+ "epochs_completed": 40,
146
+ "inference_seconds": 0.14610700006596744,
147
+ "mae_mpa": 0.00717250257730484,
148
+ "maximum_plastic_strain_trace": 1.594889909029007e-08,
149
+ "mean_consistency_relative_residual": 1.5293451838260808e-08,
150
+ "mean_peak_mises_relative_error": 8.727709791855887e-05,
151
+ "mean_work_relative_error": 7.080644718371332e-05,
152
+ "parameter_count": 3265,
153
+ "peeq_nonmonotone_fraction": 0.0,
154
+ "peeq_rmse": 2.1306934741005534e-06,
155
+ "plastic_strain_rmse": 1.7611543512430217e-07,
156
+ "r2": 0.9999999403953552,
157
+ "rmse_mpa": 0.027497082948684692,
158
+ "trajectory_count": 104
159
+ },
160
+ "physics_state_gru": {
161
+ "backstress_rmse_mpa": 24.6581974029541,
162
+ "best_validation_loss": 0.06402110308408737,
163
+ "decimated_121_state_rmse_mpa": 33.04736328125,
164
+ "epochs_completed": 45,
165
+ "inference_seconds": 0.08246395806781948,
166
+ "mae_mpa": 15.494670867919922,
167
+ "maximum_plastic_strain_trace": 0.0,
168
+ "mean_peak_mises_relative_error": 0.13843339681625366,
169
+ "mean_work_relative_error": 0.0545429065823555,
170
+ "mean_yield_surface_relative_residual_using_reference_state": 0.14034403860569,
171
+ "parameter_count": 63771,
172
+ "peeq_nonmonotone_fraction": 0.0,
173
+ "peeq_rmse": 0.004601166117936373,
174
+ "plastic_strain_rmse": 0.0001689404743956402,
175
+ "r2": 0.9557068347930908,
176
+ "rmse_mpa": 24.191770553588867,
177
+ "trajectory_count": 104
178
+ },
179
+ "pointwise_mlp": {
180
+ "best_validation_loss": 0.6130050420761108,
181
+ "decimated_121_state_rmse_mpa": 82.38230895996094,
182
+ "epochs_completed": 35,
183
+ "inference_seconds": 0.0040411248337477446,
184
+ "mae_mpa": 49.8271484375,
185
+ "mean_peak_mises_relative_error": 0.27659720182418823,
186
+ "mean_work_relative_error": 0.8062888383865356,
187
+ "mean_yield_surface_relative_residual_using_reference_state": 0.5177444815635681,
188
+ "parameter_count": 11718,
189
+ "r2": 0.4868621230125427,
190
+ "rmse_mpa": 82.34101867675781,
191
+ "trajectory_count": 104
192
+ }
193
+ },
194
+ "path_ood": {
195
+ "causal_tcn": {
196
+ "best_validation_loss": 0.5698922276496887,
197
+ "decimated_121_state_rmse_mpa": 69.38557434082031,
198
+ "epochs_completed": 17,
199
+ "inference_seconds": 0.21380204195156693,
200
+ "mae_mpa": 41.663970947265625,
201
+ "mean_peak_mises_relative_error": 0.21847069263458252,
202
+ "mean_work_relative_error": 0.21869471669197083,
203
+ "mean_yield_surface_relative_residual_using_reference_state": 0.2964002192020416,
204
+ "parameter_count": 64878,
205
+ "r2": 0.7293070554733276,
206
+ "rmse_mpa": 60.438663482666016,
207
+ "trajectory_count": 256
208
+ },
209
+ "gru": {
210
+ "best_validation_loss": 0.4674840569496155,
211
+ "decimated_121_state_rmse_mpa": 78.77181243896484,
212
+ "epochs_completed": 20,
213
+ "inference_seconds": 0.09314137510955334,
214
+ "mae_mpa": 50.5520133972168,
215
+ "mean_peak_mises_relative_error": 0.14801432192325592,
216
+ "mean_work_relative_error": 0.44539982080459595,
217
+ "mean_yield_surface_relative_residual_using_reference_state": 0.2705483138561249,
218
+ "parameter_count": 51846,
219
+ "r2": 0.5716570019721985,
220
+ "rmse_mpa": 76.02776336669922,
221
+ "trajectory_count": 256
222
+ },
223
+ "lstm": {
224
+ "best_validation_loss": 1.1083643436431885,
225
+ "decimated_121_state_rmse_mpa": 100.59011840820312,
226
+ "epochs_completed": 15,
227
+ "inference_seconds": 0.13786079199053347,
228
+ "mae_mpa": 71.41024780273438,
229
+ "mean_peak_mises_relative_error": 0.2075491100549698,
230
+ "mean_work_relative_error": 0.7572797536849976,
231
+ "mean_yield_surface_relative_residual_using_reference_state": 0.34686708450317383,
232
+ "parameter_count": 68982,
233
+ "r2": 0.2207956314086914,
234
+ "rmse_mpa": 102.54206085205078,
235
+ "trajectory_count": 256
236
+ },
237
+ "physics_integrator_nn": {
238
+ "backstress_rmse_mpa": 1.1767385331040714e-05,
239
+ "best_validation_loss": 0.00030464347219094634,
240
+ "decimated_121_state_rmse_mpa": 1.4432722330093384,
241
+ "epochs_completed": 40,
242
+ "inference_seconds": 0.17818958405405283,
243
+ "mae_mpa": 0.006917071063071489,
244
+ "maximum_plastic_strain_trace": 4.700268618762493e-09,
245
+ "mean_consistency_relative_residual": 1.490870005227407e-08,
246
+ "mean_peak_mises_relative_error": 0.00012818712275475264,
247
+ "mean_work_relative_error": 3.545045183273032e-05,
248
+ "parameter_count": 3265,
249
+ "peeq_nonmonotone_fraction": 0.0,
250
+ "peeq_rmse": 1.1404355291233514e-06,
251
+ "plastic_strain_rmse": 1.6998717455862788e-07,
252
+ "r2": 0.9999999403953552,
253
+ "rmse_mpa": 0.023686865344643593,
254
+ "trajectory_count": 256
255
+ },
256
+ "physics_state_gru": {
257
+ "backstress_rmse_mpa": 50.93850326538086,
258
+ "best_validation_loss": 0.43057119846343994,
259
+ "decimated_121_state_rmse_mpa": 84.1567611694336,
260
+ "epochs_completed": 23,
261
+ "inference_seconds": 0.13700000010430813,
262
+ "mae_mpa": 41.87034225463867,
263
+ "maximum_plastic_strain_trace": 0.0,
264
+ "mean_peak_mises_relative_error": 0.5517242550849915,
265
+ "mean_work_relative_error": 0.10701927542686462,
266
+ "mean_yield_surface_relative_residual_using_reference_state": 0.3617136776447296,
267
+ "parameter_count": 63771,
268
+ "peeq_nonmonotone_fraction": 0.0,
269
+ "peeq_rmse": 0.009458118118345737,
270
+ "plastic_strain_rmse": 0.0004539861693046987,
271
+ "r2": 0.6771982908248901,
272
+ "rmse_mpa": 66.00005340576172,
273
+ "trajectory_count": 256
274
+ },
275
+ "pointwise_mlp": {
276
+ "best_validation_loss": 1.2954227924346924,
277
+ "decimated_121_state_rmse_mpa": 105.7549057006836,
278
+ "epochs_completed": 15,
279
+ "inference_seconds": 0.009590624831616879,
280
+ "mae_mpa": 75.29700469970703,
281
+ "mean_peak_mises_relative_error": 0.29013746976852417,
282
+ "mean_work_relative_error": 0.968987226486206,
283
+ "mean_yield_surface_relative_residual_using_reference_state": 0.5085700750350952,
284
+ "parameter_count": 11718,
285
+ "r2": 0.17033731937408447,
286
+ "rmse_mpa": 105.81011199951172,
287
+ "trajectory_count": 256
288
+ }
289
+ }
290
+ }
artifacts/t2_multiaxial_ood_v2/model_protocol_comparison.png ADDED

Git LFS Details

  • SHA256: 934b3cb6c5092849141f18ae7914ad58463b4f99ce4accd3a8d8eb390fca9a95
  • Pointer size: 131 Bytes
  • Size of remote file: 102 kB
artifacts/t2_multiaxial_ood_v2/multiaxial_hysteresis_preview.png ADDED

Git LFS Details

  • SHA256: 4651c319a4d4bfab88efb42b54467619b03c9a14175e1eeffb94ad7216b25fa4
  • Pointer size: 131 Bytes
  • Size of remote file: 438 kB
artifacts/t2_multiaxial_ood_v2/physics_integrator_summary.json ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "physics_integrator_nn",
3
+ "protocols": {
4
+ "id": {
5
+ "metrics": {
6
+ "backstress_rmse_mpa": 1.1107648788311053e-05,
7
+ "best_validation_loss": 0.0002665960637386888,
8
+ "decimated_121_state_rmse_mpa": 1.0767621994018555,
9
+ "epochs_completed": 40,
10
+ "inference_seconds": 0.1447634168434888,
11
+ "mae_mpa": 0.004262915346771479,
12
+ "maximum_plastic_strain_trace": 6.51925802230835e-09,
13
+ "mean_consistency_relative_residual": 1.493731183188629e-08,
14
+ "mean_peak_mises_relative_error": 3.548976019374095e-05,
15
+ "mean_work_relative_error": 4.189212995697744e-05,
16
+ "parameter_count": 3265,
17
+ "peeq_nonmonotone_fraction": 0.0,
18
+ "peeq_rmse": 7.873447884776397e-07,
19
+ "plastic_strain_rmse": 8.518962602011015e-08,
20
+ "r2": 1.0,
21
+ "rmse_mpa": 0.012793608009815216,
22
+ "trajectory_count": 128
23
+ },
24
+ "plastic_training_transitions": 147778,
25
+ "target_log_correction_max": 0.14945420622825623,
26
+ "target_log_correction_min": -0.11471707373857498
27
+ },
28
+ "parameter_ood": {
29
+ "metrics": {
30
+ "backstress_rmse_mpa": 1.0817180736921728e-05,
31
+ "best_validation_loss": 0.00028723079594783485,
32
+ "decimated_121_state_rmse_mpa": 1.1382688283920288,
33
+ "epochs_completed": 40,
34
+ "inference_seconds": 0.14610700006596744,
35
+ "mae_mpa": 0.00717250257730484,
36
+ "maximum_plastic_strain_trace": 1.594889909029007e-08,
37
+ "mean_consistency_relative_residual": 1.5293451838260808e-08,
38
+ "mean_peak_mises_relative_error": 8.727709791855887e-05,
39
+ "mean_work_relative_error": 7.080644718371332e-05,
40
+ "parameter_count": 3265,
41
+ "peeq_nonmonotone_fraction": 0.0,
42
+ "peeq_rmse": 2.1306934741005534e-06,
43
+ "plastic_strain_rmse": 1.7611543512430217e-07,
44
+ "r2": 0.9999999403953552,
45
+ "rmse_mpa": 0.027497082948684692,
46
+ "trajectory_count": 104
47
+ },
48
+ "plastic_training_transitions": 157895,
49
+ "target_log_correction_max": 0.14945420622825623,
50
+ "target_log_correction_min": -0.11471707373857498
51
+ },
52
+ "path_ood": {
53
+ "metrics": {
54
+ "backstress_rmse_mpa": 1.1767385331040714e-05,
55
+ "best_validation_loss": 0.00030464347219094634,
56
+ "decimated_121_state_rmse_mpa": 1.4432722330093384,
57
+ "epochs_completed": 40,
58
+ "inference_seconds": 0.17818958405405283,
59
+ "mae_mpa": 0.006917071063071489,
60
+ "maximum_plastic_strain_trace": 4.700268618762493e-09,
61
+ "mean_consistency_relative_residual": 1.490870005227407e-08,
62
+ "mean_peak_mises_relative_error": 0.00012818712275475264,
63
+ "mean_work_relative_error": 3.545045183273032e-05,
64
+ "parameter_count": 3265,
65
+ "peeq_nonmonotone_fraction": 0.0,
66
+ "peeq_rmse": 1.1404355291233514e-06,
67
+ "plastic_strain_rmse": 1.6998717455862788e-07,
68
+ "r2": 0.9999999403953552,
69
+ "rmse_mpa": 0.023686865344643593,
70
+ "trajectory_count": 256
71
+ },
72
+ "plastic_training_transitions": 115735,
73
+ "target_log_correction_max": 0.13991305232048035,
74
+ "target_log_correction_min": -0.11471707373857498
75
+ }
76
+ },
77
+ "status": "completed"
78
+ }
artifacts/t2_multiaxial_ood_v2/quality.json ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "agentfem_commit": "058faecc05aeda143d014fd229401003a9258bbb",
3
+ "agentfem_version": "0.3.7.dev0",
4
+ "failures": [],
5
+ "maximum_plastic_flow_alignment_error": 6.661338147750939e-16,
6
+ "maximum_refined_peeq_relative_change": 0.004463458459340644,
7
+ "maximum_refined_stress_relative_change": 0.014655845099191836,
8
+ "maximum_rotation_relative_error": 3.6102711666568406e-15,
9
+ "maximum_yield_surface_relative_residual": 1.5751745491650184e-11,
10
+ "minimum_nonproportionality_ratio_nonproportional_families": 0.42414333383517405,
11
+ "preview": "artifacts/t2_multiaxial_ood_v2/multiaxial_hysteresis_preview.png",
12
+ "quality_failure_count": 0,
13
+ "refinement_audits": [
14
+ {
15
+ "index": 0,
16
+ "material_model": "j2_linear_isotropic",
17
+ "maximum_peeq_relative_change": 1.3581380237262815e-15,
18
+ "maximum_stress_relative_change": 7.910012430778e-16,
19
+ "path_family": "proportional_axial_cycle"
20
+ },
21
+ {
22
+ "index": 64,
23
+ "material_model": "j2_linear_isotropic",
24
+ "maximum_peeq_relative_change": 4.3337168865142253e-16,
25
+ "maximum_stress_relative_change": 3.922078645124724e-16,
26
+ "path_family": "proportional_shear_cycle"
27
+ },
28
+ {
29
+ "index": 128,
30
+ "material_model": "j2_linear_isotropic",
31
+ "maximum_peeq_relative_change": 0.0012345909704224037,
32
+ "maximum_stress_relative_change": 0.014655845099191836,
33
+ "path_family": "sequential_axial_shear"
34
+ },
35
+ {
36
+ "index": 192,
37
+ "material_model": "j2_linear_isotropic",
38
+ "maximum_peeq_relative_change": 0.0008065711576948378,
39
+ "maximum_stress_relative_change": 0.006567542855441546,
40
+ "path_family": "orthogonal_cross"
41
+ },
42
+ {
43
+ "index": 256,
44
+ "material_model": "j2_linear_isotropic",
45
+ "maximum_peeq_relative_change": 0.0017817981166268794,
46
+ "maximum_stress_relative_change": 0.009095677551799521,
47
+ "path_family": "nonproportional_square"
48
+ },
49
+ {
50
+ "index": 320,
51
+ "material_model": "j2_linear_isotropic",
52
+ "maximum_peeq_relative_change": 0.002359786008358582,
53
+ "maximum_stress_relative_change": 0.008340815204171284,
54
+ "path_family": "rotating_principal_cycle"
55
+ },
56
+ {
57
+ "index": 384,
58
+ "material_model": "j2_linear_isotropic",
59
+ "maximum_peeq_relative_change": 0.004463458459340644,
60
+ "maximum_stress_relative_change": 0.011535422602253764,
61
+ "path_family": "out_of_phase_lissajous"
62
+ },
63
+ {
64
+ "index": 448,
65
+ "material_model": "j2_linear_isotropic",
66
+ "maximum_peeq_relative_change": 0.002574167101098269,
67
+ "maximum_stress_relative_change": 0.010450297669248865,
68
+ "path_family": "random_direction_blocks"
69
+ },
70
+ {
71
+ "index": 512,
72
+ "material_model": "chaboche_combined",
73
+ "maximum_peeq_relative_change": 0.00012206957356614427,
74
+ "maximum_stress_relative_change": 0.00012625650150541095,
75
+ "path_family": "proportional_axial_cycle"
76
+ },
77
+ {
78
+ "index": 576,
79
+ "material_model": "chaboche_combined",
80
+ "maximum_peeq_relative_change": 0.0003478735659638992,
81
+ "maximum_stress_relative_change": 0.0009797651936342881,
82
+ "path_family": "proportional_shear_cycle"
83
+ },
84
+ {
85
+ "index": 640,
86
+ "material_model": "chaboche_combined",
87
+ "maximum_peeq_relative_change": 0.0024305653920597753,
88
+ "maximum_stress_relative_change": 0.0038191269609130314,
89
+ "path_family": "sequential_axial_shear"
90
+ },
91
+ {
92
+ "index": 704,
93
+ "material_model": "chaboche_combined",
94
+ "maximum_peeq_relative_change": 0.00023802202906946098,
95
+ "maximum_stress_relative_change": 0.008407205515433933,
96
+ "path_family": "orthogonal_cross"
97
+ },
98
+ {
99
+ "index": 768,
100
+ "material_model": "chaboche_combined",
101
+ "maximum_peeq_relative_change": 0.0017385285228106618,
102
+ "maximum_stress_relative_change": 0.008219242610358803,
103
+ "path_family": "nonproportional_square"
104
+ },
105
+ {
106
+ "index": 832,
107
+ "material_model": "chaboche_combined",
108
+ "maximum_peeq_relative_change": 0.001927979198753472,
109
+ "maximum_stress_relative_change": 0.004268020616286278,
110
+ "path_family": "rotating_principal_cycle"
111
+ },
112
+ {
113
+ "index": 896,
114
+ "material_model": "chaboche_combined",
115
+ "maximum_peeq_relative_change": 0.0020326086374519736,
116
+ "maximum_stress_relative_change": 0.004881998122870186,
117
+ "path_family": "out_of_phase_lissajous"
118
+ },
119
+ {
120
+ "index": 960,
121
+ "material_model": "chaboche_combined",
122
+ "maximum_peeq_relative_change": 0.0015114962366540713,
123
+ "maximum_stress_relative_change": 0.007169790637155542,
124
+ "path_family": "random_direction_blocks"
125
+ }
126
+ ],
127
+ "rotation_audits": [
128
+ {
129
+ "index": 0,
130
+ "plastic_strain_relative_error": 2.385366890650902e-16,
131
+ "stress_relative_error": 2.7916404349832403e-15
132
+ },
133
+ {
134
+ "index": 64,
135
+ "plastic_strain_relative_error": 2.2929428389957014e-16,
136
+ "stress_relative_error": 2.616238013145273e-16
137
+ },
138
+ {
139
+ "index": 128,
140
+ "plastic_strain_relative_error": 2.887125803370766e-16,
141
+ "stress_relative_error": 3.6102711666568406e-15
142
+ },
143
+ {
144
+ "index": 192,
145
+ "plastic_strain_relative_error": 2.6808884553327007e-16,
146
+ "stress_relative_error": 4.3526651263636753e-16
147
+ },
148
+ {
149
+ "index": 256,
150
+ "plastic_strain_relative_error": 2.2685340943466496e-16,
151
+ "stress_relative_error": 1.097691556523258e-15
152
+ },
153
+ {
154
+ "index": 320,
155
+ "plastic_strain_relative_error": 2.8143975854116957e-16,
156
+ "stress_relative_error": 2.8180292203525067e-15
157
+ },
158
+ {
159
+ "index": 384,
160
+ "plastic_strain_relative_error": 2.2673408284763576e-16,
161
+ "stress_relative_error": 8.837288127389734e-16
162
+ },
163
+ {
164
+ "index": 448,
165
+ "plastic_strain_relative_error": 2.913979641560205e-16,
166
+ "stress_relative_error": 1.720975259738966e-15
167
+ },
168
+ {
169
+ "index": 512,
170
+ "plastic_strain_relative_error": 2.459658188792725e-16,
171
+ "stress_relative_error": 3.7062352591722016e-16
172
+ },
173
+ {
174
+ "index": 576,
175
+ "plastic_strain_relative_error": 2.0778767248203813e-16,
176
+ "stress_relative_error": 7.566969568673829e-16
177
+ },
178
+ {
179
+ "index": 640,
180
+ "plastic_strain_relative_error": 2.8463764198356526e-16,
181
+ "stress_relative_error": 3.714014716108652e-16
182
+ },
183
+ {
184
+ "index": 704,
185
+ "plastic_strain_relative_error": 2.839414670125401e-16,
186
+ "stress_relative_error": 2.594130842090602e-15
187
+ },
188
+ {
189
+ "index": 768,
190
+ "plastic_strain_relative_error": 2.1836469171385463e-16,
191
+ "stress_relative_error": 8.838305285714192e-16
192
+ },
193
+ {
194
+ "index": 832,
195
+ "plastic_strain_relative_error": 3.8215614291928256e-16,
196
+ "stress_relative_error": 1.713333619543387e-15
197
+ },
198
+ {
199
+ "index": 896,
200
+ "plastic_strain_relative_error": 4.453473217004695e-16,
201
+ "stress_relative_error": 1.4799263773448094e-15
202
+ },
203
+ {
204
+ "index": 960,
205
+ "plastic_strain_relative_error": 2.797071129690676e-16,
206
+ "stress_relative_error": 6.654281173648105e-16
207
+ }
208
+ ],
209
+ "sample_count": 1024,
210
+ "split_counts": {
211
+ "split_id": {
212
+ "test": 128,
213
+ "train": 768,
214
+ "validation": 128
215
+ },
216
+ "split_parameter_ood": {
217
+ "test": 104,
218
+ "train": 818,
219
+ "validation": 102
220
+ },
221
+ "split_path_ood": {
222
+ "test": 256,
223
+ "train": 640,
224
+ "validation": 128
225
+ }
226
+ },
227
+ "status": "accepted"
228
+ }
artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/notched_bar_comparison.png ADDED

Git LFS Details

  • SHA256: 69b803b9a1449bda3e8208fe27bcc1e868e4408b8582c2fb9c56a128922e971c
  • Pointer size: 131 Bytes
  • Size of remote file: 328 kB
artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/notched_bar_results.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8eb1650c1358f618f02bd51e668c23dfe57f12555a9f8c8fa97df25f3c9cd889
3
+ size 41458
artifacts/t2_multiaxial_ood_v2/structural_validation/mild_id/structural_validation.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "case_name": "mild_id",
3
+ "elements": 12,
4
+ "limitations": [
5
+ "This is a reduced one-dimensional structural gate, not a general 3D learned AgentFEM material provider.",
6
+ "The physics-integrator uses a numerical tangent of its fully discrete update and records any trust-region fallback used when plain global Newton leaves its local basin.",
7
+ "The loading is proportional at each material point even though the training dataset is multiaxial."
8
+ ],
9
+ "load_points": 81,
10
+ "load_scale": 0.65,
11
+ "material_model": "chaboche_combined",
12
+ "model_failures": {},
13
+ "models": {
14
+ "gru": {
15
+ "elapsed_seconds": 13.867982124909759,
16
+ "maximum_newton_iterations": 5,
17
+ "mean_newton_iterations": 2.876543209876543,
18
+ "reaction_max_relative_error": 0.384637991446209,
19
+ "reaction_relative_l2": 0.255284459318306,
20
+ "strain_relative_l2": 0.10129071845746153,
21
+ "stress_max_relative_error": 0.3846380164058209,
22
+ "stress_relative_l2": 0.2552844280184351
23
+ },
24
+ "physics_integrator_nn": {
25
+ "elapsed_seconds": 10.280943374847993,
26
+ "maximum_newton_iterations": 5,
27
+ "mean_newton_iterations": 2.5555555555555554,
28
+ "reaction_max_relative_error": 6.141174169837029e-07,
29
+ "reaction_relative_l2": 1.6513475327397011e-07,
30
+ "strain_relative_l2": 1.655642344026917e-06,
31
+ "stress_max_relative_error": 1.4670657579488333e-06,
32
+ "stress_relative_l2": 2.504266086571615e-07,
33
+ "trust_region_fallback_count": 0,
34
+ "trust_region_fallback_steps": []
35
+ },
36
+ "physics_state_gru": {
37
+ "elapsed_seconds": 13.064999458845705,
38
+ "maximum_newton_iterations": 4,
39
+ "mean_newton_iterations": 2.7037037037037037,
40
+ "reaction_max_relative_error": 0.7323474985512538,
41
+ "reaction_relative_l2": 0.5238098222184954,
42
+ "strain_relative_l2": 0.12406144632033132,
43
+ "stress_max_relative_error": 0.7323475275605655,
44
+ "stress_relative_l2": 0.5238098490729751
45
+ }
46
+ },
47
+ "native_elapsed_seconds": 18.18587166699581,
48
+ "notch_depth_fraction": 0.12,
49
+ "scope": "controlled one-dimensional notched-bar finite-element deployment",
50
+ "status": "completed"
51
+ }
artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/notched_bar_comparison.png ADDED

Git LFS Details

  • SHA256: d7af84c60b140016d5abba35454a63d84bd58584911d0496760d35b9f276c01f
  • Pointer size: 131 Bytes
  • Size of remote file: 279 kB
artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/notched_bar_results.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ded38f99016820dacb06d1a325b7cf6fd13a1d7053020803ff6c45ec2c2ec268
3
+ size 30569
artifacts/t2_multiaxial_ood_v2/structural_validation/moderate_ood/structural_validation.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "case_name": "moderate_ood",
3
+ "elements": 12,
4
+ "limitations": [
5
+ "This is a reduced one-dimensional structural gate, not a general 3D learned AgentFEM material provider.",
6
+ "Learned tangents are differentiated by autograd and clipped for global robustness; clipping events are not yet exposed.",
7
+ "The loading is proportional at each material point even though the training dataset is multiaxial."
8
+ ],
9
+ "load_points": 81,
10
+ "load_scale": 0.85,
11
+ "material_model": "chaboche_combined",
12
+ "model_failures": {
13
+ "physics_integrator_nn": "physics_integrator_nn structural solve failed at step 58; max_abs_element_strain=0.000480067."
14
+ },
15
+ "models": {
16
+ "gru": {
17
+ "elapsed_seconds": 15.302196583943442,
18
+ "maximum_newton_iterations": 14,
19
+ "mean_newton_iterations": 3.234567901234568,
20
+ "reaction_max_relative_error": 0.4188433021883623,
21
+ "reaction_relative_l2": 0.317760889891088,
22
+ "strain_relative_l2": 0.19402064840808173,
23
+ "stress_max_relative_error": 0.41884318632562195,
24
+ "stress_relative_l2": 0.3177609297607427
25
+ },
26
+ "physics_state_gru": {
27
+ "elapsed_seconds": 17.36364437500015,
28
+ "maximum_newton_iterations": 22,
29
+ "mean_newton_iterations": 3.3209876543209877,
30
+ "reaction_max_relative_error": 0.929228474435714,
31
+ "reaction_relative_l2": 0.5957133727527093,
32
+ "strain_relative_l2": 0.29032182174573423,
33
+ "stress_max_relative_error": 0.9292284577775981,
34
+ "stress_relative_l2": 0.5957132741048787
35
+ }
36
+ },
37
+ "native_elapsed_seconds": 20.369384250137955,
38
+ "notch_depth_fraction": 0.3,
39
+ "scope": "controlled one-dimensional notched-bar finite-element deployment",
40
+ "status": "completed_with_model_failures"
41
+ }
artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/notched_bar_comparison.png ADDED

Git LFS Details

  • SHA256: 740e645838d4b3276169a72e304aee7cab054dab8a63edb60d5867a87d63af57
  • Pointer size: 131 Bytes
  • Size of remote file: 290 kB
artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/notched_bar_results.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:999ef1d758cedec3d918662f223448826ff037b86423821f99d7f3e8bb959d51
3
+ size 33057
artifacts/t2_multiaxial_ood_v2/structural_validation/severe_ood/structural_validation.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "case_name": "severe_ood",
3
+ "elements": 12,
4
+ "limitations": [
5
+ "This is a reduced one-dimensional structural gate, not a general 3D learned AgentFEM material provider.",
6
+ "The physics-integrator uses a numerical tangent of its fully discrete update and records any trust-region fallback used when plain global Newton leaves its local basin.",
7
+ "The loading is proportional at each material point even though the training dataset is multiaxial."
8
+ ],
9
+ "load_points": 81,
10
+ "load_scale": 1.0,
11
+ "material_model": "chaboche_combined",
12
+ "model_failures": {
13
+ "physics_state_gru": "physics_state_gru structural solve failed at step 58."
14
+ },
15
+ "models": {
16
+ "gru": {
17
+ "elapsed_seconds": 17.71098691690713,
18
+ "maximum_newton_iterations": 25,
19
+ "mean_newton_iterations": 3.6419753086419755,
20
+ "reaction_max_relative_error": 0.4814582275953355,
21
+ "reaction_relative_l2": 0.3703393789109682,
22
+ "strain_relative_l2": 0.22436863535185508,
23
+ "stress_max_relative_error": 0.4814580924697208,
24
+ "stress_relative_l2": 0.3703391386706923
25
+ },
26
+ "physics_integrator_nn": {
27
+ "elapsed_seconds": 23.848188207950443,
28
+ "maximum_newton_iterations": 48,
29
+ "mean_newton_iterations": 3.9382716049382718,
30
+ "reaction_max_relative_error": 8.784767139075504e-07,
31
+ "reaction_relative_l2": 2.3921409855059484e-07,
32
+ "strain_relative_l2": 2.0695867029594525e-07,
33
+ "stress_max_relative_error": 6.546060134598156e-07,
34
+ "stress_relative_l2": 2.0073940000059515e-07,
35
+ "trust_region_fallback_count": 1,
36
+ "trust_region_fallback_steps": [
37
+ 58
38
+ ]
39
+ }
40
+ },
41
+ "native_elapsed_seconds": 24.727217457955703,
42
+ "notch_depth_fraction": 0.42,
43
+ "scope": "controlled one-dimensional notched-bar finite-element deployment",
44
+ "status": "completed_with_model_failures"
45
+ }
artifacts/t2_multiaxial_ood_v2/training_summary.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "batch_size": 32,
3
+ "dataset_sample_count": 1024,
4
+ "elapsed_seconds": 331.024352916982,
5
+ "epochs_requested": 45,
6
+ "metrics_path": "artifacts/t2_multiaxial_ood_v2/model_metrics.json",
7
+ "model_directory": "models/t2_multiaxial_ood_v2",
8
+ "models": [
9
+ "physics_state_gru"
10
+ ],
11
+ "protocols": [
12
+ "id",
13
+ "path_ood",
14
+ "parameter_ood"
15
+ ],
16
+ "seed": 20261011,
17
+ "status": "completed"
18
+ }
configs/t2_multiaxial_ood_v2.json ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "agentfem.physics-data.multiaxial-constitutive-memory",
3
+ "schema_version": "2.0.0",
4
+ "dataset_version": "2.0.0-research",
5
+ "seed": 20261001,
6
+ "sample_count": 1024,
7
+ "samples_per_stratum": 64,
8
+ "points_per_trajectory": 241,
9
+ "shard_count": 8,
10
+ "material_models": {
11
+ "j2_linear_isotropic": 512,
12
+ "chaboche_combined": 512
13
+ },
14
+ "path_families": {
15
+ "proportional_axial_cycle": 128,
16
+ "proportional_shear_cycle": 128,
17
+ "sequential_axial_shear": 128,
18
+ "orthogonal_cross": 128,
19
+ "nonproportional_square": 128,
20
+ "rotating_principal_cycle": 128,
21
+ "out_of_phase_lissajous": 128,
22
+ "random_direction_blocks": 128
23
+ },
24
+ "id_splits_per_stratum": {
25
+ "train": 48,
26
+ "validation": 8,
27
+ "test": 8
28
+ },
29
+ "path_ood_families": {
30
+ "train": [
31
+ "proportional_axial_cycle",
32
+ "proportional_shear_cycle",
33
+ "sequential_axial_shear",
34
+ "orthogonal_cross",
35
+ "nonproportional_square"
36
+ ],
37
+ "validation": ["rotating_principal_cycle"],
38
+ "test": ["out_of_phase_lissajous", "random_direction_blocks"]
39
+ },
40
+ "ranges": {
41
+ "young_pa": [150000000000.0, 230000000000.0],
42
+ "poisson": [0.25, 0.34],
43
+ "yield_stress_pa": [180000000.0, 450000000.0],
44
+ "maximum_equivalent_strain": [0.003, 0.012],
45
+ "j2_hardening_modulus_pa": [500000000.0, 8000000000.0],
46
+ "chaboche_c1_pa": [15000000000.0, 60000000000.0],
47
+ "chaboche_gamma1": [20.0, 120.0],
48
+ "chaboche_c2_pa": [2000000000.0, 15000000000.0],
49
+ "chaboche_gamma2": [1.0, 20.0],
50
+ "chaboche_isotropic_saturation_pa": [20000000.0, 120000000.0],
51
+ "chaboche_isotropic_rate": [2.0, 20.0]
52
+ },
53
+ "quality_thresholds": {
54
+ "initial_stress_pa": 0.001,
55
+ "stress_symmetry_pa": 0.001,
56
+ "plastic_strain_trace": 1e-10,
57
+ "equivalent_plastic_strain_decrease": 1e-12,
58
+ "yield_surface_relative_residual": 1e-8,
59
+ "plastic_flow_alignment_error": 1e-8,
60
+ "rotation_relative_error": 1e-9,
61
+ "refined_stress_relative_change": 0.03,
62
+ "refined_peeq_relative_change": 0.06
63
+ },
64
+ "scope": "Synthetic three-dimensional small-strain material-point histories under proportional and non-proportional deviatoric strain. J2 linear isotropic hardening and experimental Chaboche combined hardening remain separate labels. No experimental calibration, fatigue-life, damage, temperature, finite-strain or structural-field claim."
65
+ }
data/t2_multiaxial_ood_v2/README.md ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentFEM Multiaxial Constitutive Memory v2 (research release)
2
+
3
+ This extension contains 1,024 verified tensor-valued material-point histories
4
+ for studying path-dependent neural constitutive models under controlled
5
+ in-distribution and out-of-distribution evaluation. It preserves the published
6
+ T2 v1 data and adds true non-proportional loading in the complete five-
7
+ dimensional deviatoric strain space.
8
+
9
+ ## Contents
10
+
11
+ - 512 J2 linear-isotropic-hardening trajectories.
12
+ - 512 Chaboche combined-hardening trajectories.
13
+ - Eight path families, including proportional cycles, sequential orthogonal
14
+ loading, five-direction crosses, rotating principal directions, Lissajous
15
+ paths and random five-dimensional direction blocks.
16
+ - 241 ordered states per trajectory. States are never counted as independent
17
+ samples.
18
+ - Full strain, stress, plastic-strain and backstress tensors; equivalent
19
+ plastic strain, plastic increments, yield data and work diagnostics.
20
+ - Independent labels for ID, path-OOD and parameter-OOD protocols.
21
+ - Six trained models for every protocol: pointwise MLP, GRU, LSTM, causal TCN,
22
+ a physics-state GRU and a physics-embedded neural stress integrator.
23
+
24
+ ## Quality evidence
25
+
26
+ All 1,024 trajectories passed the release gates.
27
+
28
+ | Check | Worst observed value |
29
+ |---|---:|
30
+ | Yield-surface relative residual | 1.575e-11 |
31
+ | Associated-flow alignment error | 6.661e-16 |
32
+ | Rigid-rotation covariance error | 3.610e-15 |
33
+ | 241-to-481-state stress change | 1.466% |
34
+ | 241-to-481-state PEEQ change | 0.446% |
35
+
36
+ The complete evidence is in
37
+ `artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md`.
38
+
39
+ ## Benchmark results
40
+
41
+ Stress RMSE on each protocol's held-out trajectories:
42
+
43
+ | Model | ID | Path OOD | Parameter OOD |
44
+ |---|---:|---:|---:|
45
+ | Pointwise MLP | 83.51 MPa | 105.81 MPa | 82.34 MPa |
46
+ | GRU | **17.26 MPa** | 76.03 MPa | **19.26 MPa** |
47
+ | LSTM | 22.67 MPa | 102.54 MPa | 22.89 MPa |
48
+ | Causal TCN | 21.82 MPa | **60.44 MPa** | 21.99 MPa |
49
+ | Physics-state GRU | 23.06 MPa | 66.00 MPa | 24.19 MPa |
50
+ | Physics-integrator NN | **0.0128 MPa** | **0.0237 MPa** | **0.0275 MPa** |
51
+
52
+ The physics-state GRU predicts plastic strain, backstress and equivalent
53
+ plastic strain. Plastic strain is exactly deviatoric, PEEQ is exactly
54
+ nondecreasing, and stress is reconstructed from elastic strain and the
55
+ isotropic elastic law. Those constraints improve admissibility but do not
56
+ guarantee unseen-path accuracy or structural-solver stability. The
57
+ physics-integrator instead places the neural correction inside a J2/Chaboche
58
+ return-mapping update and explicitly evolves the internal variables. Its very
59
+ small errors are a white-box ceiling because the embedded equations match the
60
+ synthetic data generator.
61
+
62
+ ![Protocol comparison](../../artifacts/t2_multiaxial_ood_v2/model_protocol_comparison.png)
63
+
64
+ ## Structural deployment gate
65
+
66
+ The ID-trained GRU, physics-state GRU and physics-integrator NN were embedded as element-level
67
+ material updates in a controlled displacement-driven notched-bar finite-
68
+ element solver and compared with native AgentFEM Chaboche updates.
69
+
70
+ - Mild 12% notch: both networks converged, but reaction relative-L2 errors
71
+ were 25.5% (GRU) and 52.4% (physics-state GRU).
72
+ - Severe 42% notch: the GRU converged with 37.0% reaction relative-L2 error;
73
+ the physics-state GRU failed global convergence at load step 58; the
74
+ physics-integrator NN converged with 2.39e-7 error after one recorded
75
+ trust-region fallback at complete unloading.
76
+
77
+ These are deliberately retained negative results. Material-point random-split
78
+ accuracy is not evidence of reliable FE deployment. The structural gate is
79
+ one-dimensional and does not claim a production learned-material provider.
80
+
81
+ ## Splits
82
+
83
+ - `split_id`: 768/128/128 train/validation/test trajectories, balanced within
84
+ each material/path stratum.
85
+ - `split_path_ood`: five path families for training, rotating-principal paths
86
+ for validation, and Lissajous plus random-direction paths for testing.
87
+ - `split_parameter_ood`: central 80% of the normalized initial-yield-stress
88
+ range for training, adjacent 5% bands for validation and outer 5% tails for
89
+ testing.
90
+ - Resolution sensitivity: trained 241-state models are also evaluated after
91
+ sequence decimation to 121 states.
92
+
93
+ These protocols answer different questions and must not be merged into a
94
+ single score.
95
+
96
+ ## Loading and units
97
+
98
+ The Voigt order is `xx, yy, zz, xy, yz, xz`; shear entries are tensor shear,
99
+ not engineering shear. All quantities use SI units. The five basis tensors are
100
+ orthonormal under `sqrt(2/3 e_dev:e_dev)`, so the Euclidean norm of a stored
101
+ path-coordinate vector equals equivalent deviatoric strain.
102
+
103
+ ## Reproduction
104
+
105
+ Start with `T2_V2_START_HERE.md` and `T2_V2_REPRODUCE.md` in the repository
106
+ root. The exact AgentFEM source revision is
107
+ `058faecc05aeda143d014fd229401003a9258bbb` (`0.3.7.dev0`).
108
+
109
+ ## Scope
110
+
111
+ The release is synthetic small-strain material-point data generated from two
112
+ classical constitutive families. It is not experimental calibration, fatigue
113
+ life, damage, finite-strain plasticity or a structural field dataset. Chaboche
114
+ remains explicitly labelled as an experimental AgentFEM capability.
data/t2_multiaxial_ood_v2/design.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/t2_multiaxial_ood_v2/index.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/t2_multiaxial_ood_v2/manifest.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "agentfem_commit": "058faecc05aeda143d014fd229401003a9258bbb",
3
+ "agentfem_version": "0.3.7.dev0",
4
+ "dataset_version": "2.0.0-research",
5
+ "points_per_trajectory": 241,
6
+ "sample_count": 1024,
7
+ "schema": "agentfem.physics-data.multiaxial-constitutive-memory",
8
+ "schema_version": "2.0.0",
9
+ "shards": [
10
+ {
11
+ "bytes": 14028846,
12
+ "first_id": "00000",
13
+ "last_id": "00127",
14
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00000.h5",
15
+ "sample_count": 128,
16
+ "sha256": "16456ca1050a8afff4b687610d0af7b952dcd2b82149ad61a13e7f1296daa1d4"
17
+ },
18
+ {
19
+ "bytes": 17135531,
20
+ "first_id": "00128",
21
+ "last_id": "00255",
22
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00001.h5",
23
+ "sample_count": 128,
24
+ "sha256": "3d56b126367a7cc2e7c7adeca5ae7badaafc91ad4130d5cab444a18c81ec3b10"
25
+ },
26
+ {
27
+ "bytes": 17561117,
28
+ "first_id": "00256",
29
+ "last_id": "00383",
30
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00002.h5",
31
+ "sample_count": 128,
32
+ "sha256": "7742b2533ae3a7c2f5018e83abfa44c236caed8c11766d357f56dc92b136f84c"
33
+ },
34
+ {
35
+ "bytes": 19857537,
36
+ "first_id": "00384",
37
+ "last_id": "00511",
38
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00003.h5",
39
+ "sample_count": 128,
40
+ "sha256": "189d2ef804c3639fb85a58bddacc982e2d7634add25f0ea214c2652f69193aa6"
41
+ },
42
+ {
43
+ "bytes": 15991512,
44
+ "first_id": "00512",
45
+ "last_id": "00639",
46
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00004.h5",
47
+ "sample_count": 128,
48
+ "sha256": "364cdd35f7d8280992267ceda7c4313a197811f8df408763c1f60a144a2a0c59"
49
+ },
50
+ {
51
+ "bytes": 21022139,
52
+ "first_id": "00640",
53
+ "last_id": "00767",
54
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00005.h5",
55
+ "sample_count": 128,
56
+ "sha256": "bf3abef7c1a1c1686725014f5be2b49189bca023014d889d418c8950a7699e80"
57
+ },
58
+ {
59
+ "bytes": 21291993,
60
+ "first_id": "00768",
61
+ "last_id": "00895",
62
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00006.h5",
63
+ "sample_count": 128,
64
+ "sha256": "27fd47d03fff7ca8e394be61d4a4af2366de9d38ed695e47cbb2ae1fa8c90af7"
65
+ },
66
+ {
67
+ "bytes": 24766067,
68
+ "first_id": "00896",
69
+ "last_id": "01023",
70
+ "path": "data/t2_multiaxial_ood_v2/shards/part-00007.h5",
71
+ "sample_count": 128,
72
+ "sha256": "9d1b9d2bc4770cd67516b2f56ba56ae951d6d072a0f0286efccd18c6e94a686e"
73
+ }
74
+ ],
75
+ "strain_shear_convention": "tensor",
76
+ "voigt_order": [
77
+ "xx",
78
+ "yy",
79
+ "zz",
80
+ "xy",
81
+ "yz",
82
+ "xz"
83
+ ]
84
+ }
data/t2_multiaxial_ood_v2/shards/part-00000.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16456ca1050a8afff4b687610d0af7b952dcd2b82149ad61a13e7f1296daa1d4
3
+ size 14028846
data/t2_multiaxial_ood_v2/shards/part-00001.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3d56b126367a7cc2e7c7adeca5ae7badaafc91ad4130d5cab444a18c81ec3b10
3
+ size 17135531
data/t2_multiaxial_ood_v2/shards/part-00002.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7742b2533ae3a7c2f5018e83abfa44c236caed8c11766d357f56dc92b136f84c
3
+ size 17561117
data/t2_multiaxial_ood_v2/shards/part-00003.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:189d2ef804c3639fb85a58bddacc982e2d7634add25f0ea214c2652f69193aa6
3
+ size 19857537
data/t2_multiaxial_ood_v2/shards/part-00004.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:364cdd35f7d8280992267ceda7c4313a197811f8df408763c1f60a144a2a0c59
3
+ size 15991512
data/t2_multiaxial_ood_v2/shards/part-00005.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bf3abef7c1a1c1686725014f5be2b49189bca023014d889d418c8950a7699e80
3
+ size 21022139
data/t2_multiaxial_ood_v2/shards/part-00006.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:27fd47d03fff7ca8e394be61d4a4af2366de9d38ed695e47cbb2ae1fa8c90af7
3
+ size 21291993
data/t2_multiaxial_ood_v2/shards/part-00007.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9d1b9d2bc4770cd67516b2f56ba56ae951d6d072a0f0286efccd18c6e94a686e
3
+ size 24766067
docs/T2_MULTIAXIAL_LITERATURE.md ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # T2 multiaxial constitutive-learning literature map
2
+
3
+ This short map is intended for manuscript preparation, not as an exhaustive
4
+ review.
5
+
6
+ 1. Gorji et al., *On the potential of recurrent neural networks for modeling
7
+ path dependent plasticity*, Journal of the Mechanics and Physics of Solids.
8
+ Establishes RNNs as substitutes for path-dependent plasticity models and
9
+ discusses loading-path design.
10
+ <https://www.sciencedirect.com/science/article/pii/S0022509620302076>
11
+
12
+ 2. Rocha et al., *Physically recurrent neural networks for path-dependent
13
+ heterogeneous materials: Embedding constitutive models in a data-driven
14
+ surrogate*, CMAME 407 (2023) 115934. Embeds physics-based material models
15
+ inside a recurrent surrogate and reports large FE2 speed-ups.
16
+ <https://doi.org/10.1016/j.cma.2023.115934>
17
+
18
+ 3. Bonatti and Mohr, *On the importance of self-consistency in recurrent neural
19
+ network models representing elasto-plastic solids*. The key connection for
20
+ this project is that a constitutive response should not depend on arbitrary
21
+ loading-increment size.
22
+ <https://www.sciencedirect.com/science/article/pii/S0022509621003161>
23
+
24
+ 4. Hashash et al., *Machine learning based modeling of path-dependent materials
25
+ for finite element analysis*, Computers and Geotechnics 156 (2023) 105254.
26
+ Directly identifies increment-size dependence as a source of FE error and
27
+ non-convergence and validates on boundary-value problems.
28
+ <https://doi.org/10.1016/j.compgeo.2023.105254>
29
+
30
+ 5. Ma et al., *Learning Neural Constitutive Laws from Motion Observations for
31
+ Generalizable PDE Dynamics*, ICML 2023. Uses strict constitutive priors and
32
+ differentiable simulation to target geometry and boundary-condition OOD
33
+ generalization.
34
+ <https://proceedings.mlr.press/v202/ma23a.html>
35
+
36
+ 6. Rezaei et al., *Learning solution of nonlinear constitutive material models
37
+ using physics-informed neural networks: COMM-PINN*. Learns internal state
38
+ evolution while enforcing constitutive equations and thermodynamic
39
+ restrictions.
40
+ <https://arxiv.org/abs/2304.06044>
41
+
42
+ 7. Meyer and Ekre, *Thermodynamically consistent neural network plasticity
43
+ modeling and discovery of evolution laws*, JMPS 2023. Embeds a neural
44
+ evolution law inside a standard dissipative constitutive structure and
45
+ enforces thermodynamics and frame invariance by construction.
46
+ <https://doi.org/10.1016/j.jmps.2023.105416>
47
+
48
+ 8. Masi et al., *Thermodynamics-based Artificial Neural Networks for
49
+ constitutive modeling*. Derives stress and dissipation from learned energy
50
+ and internal-state representations, providing the thermodynamic reference
51
+ point for latent-variable extensions.
52
+ <https://arxiv.org/abs/2005.12183>
53
+
54
+ 9. Ghavamian and Simone, *Machine learning of evolving physics-based material
55
+ models for multiscale solid mechanics*, Mechanics of Materials 184 (2023)
56
+ 104707. Uses a physics decoder with learned evolving parameters, a direct
57
+ precedent for a gray-box extension beyond the present white-box benchmark.
58
+ <https://doi.org/10.1016/j.mechmat.2023.104707>
59
+
60
+ ## Gap addressed by the present release
61
+
62
+ Many studies demonstrate material-point accuracy or selected successful FE
63
+ examples. This release instead freezes several distinct OOD protocols, reports
64
+ physical-state violations, changes temporal resolution, retains failed
65
+ structural deployment, and adds a neural-seeded return-mapping model. Its main
66
+ research value is the evidence chain between material-point prediction,
67
+ physical admissibility, time-increment sensitivity and global-solver behavior.
docs/T2_MULTIAXIAL_RESEARCH_MEMO.md ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # T2多轴本构记忆研究备忘录
2
+
3
+ ## 一、建议文章主线
4
+
5
+ 建议题目方向:
6
+
7
+ **从材料点高精度到结构求解失效:路径相关神经本构模型的多轴OOD基准与有限元部署验证**
8
+
9
+ 核心问题不是比较“哪个网络分数最高”,而是研究:随机留出测试中的高精度,能否迁移到未知加载路径、材料参数外推、时间离散变化和结构有限元全局平衡。
10
+
11
+ 当前结果给出了清晰答案:材料参数外推相对容易,未知路径明显更难;仅约束输出状态的网络不能自动保证OOD精度和全局Newton稳定性。将神经网络嵌入增量塑性积分,并保留返回映射、内部变量和屈服一致性后,材料点OOD和结构部署精度均显著提高。
12
+
13
+ ## 二、已经完成的证据
14
+
15
+ ### 数据与物理验证
16
+
17
+ - 1,024条独立轨迹,J2与Chaboche各512条;
18
+ - 每条241个有序状态点,覆盖完整五维偏应变空间;
19
+ - 8类加载路径,其中6类为非比例多轴路径;
20
+ - ID、路径OOD、参数OOD三套独立划分;
21
+ - 屈服面最大相对残差1.575×10⁻¹¹;
22
+ - 关联流动最大方向误差6.661×10⁻¹⁶;
23
+ - 旋转协变最大误差3.610×10⁻¹⁵;
24
+ - 241→481点加密最大应力差1.466%,PEEQ差0.446%;
25
+ - 0个质量门失败样本。
26
+
27
+ ### 神经网络比较
28
+
29
+ | 模型 | ID RMSE | 路径OOD RMSE | 参数OOD RMSE |
30
+ |---|---:|---:|---:|
31
+ | 点式MLP | 83.51 MPa | 105.81 MPa | 82.34 MPa |
32
+ | GRU | 17.26 MPa | 76.03 MPa | 19.26 MPa |
33
+ | LSTM | 22.67 MPa | 102.54 MPa | 22.89 MPa |
34
+ | 因果TCN | 21.82 MPa | 60.44 MPa | 21.99 MPa |
35
+ | 物理状态GRU | 23.06 MPa | 66.00 MPa | 24.19 MPa |
36
+ | 物理积分网络 | **0.0128 MPa** | **0.0237 MPa** | **0.0275 MPa** |
37
+
38
+ 最重要的结果是:GRU在ID下R²达到0.978,但路径OOD误差增至76.03 MPa。参数OOD仍能保持19.26 MPa,说明“未知材料参数”和“未知路径拓扑”不是同一种泛化问题。
39
+
40
+ ### 物理状态模型
41
+
42
+ 物理状态GRU不直接预测应力,而是预测塑性应变、背应力和PEEQ:
43
+
44
+ 1. 用五个独立分量构造无迹塑性应变,严格满足塑性不可压缩;
45
+ 2. 用非负增量累加PEEQ,严格保证单调不减;
46
+ 3. 由总应变减塑性应变得到弹性应变,再用各向同性弹性关系重构应力;
47
+ 4. 训练损失包含屈服面、内部变量与应力误差。
48
+
49
+ 该模型在ID测试中应力RMSE为23.06 MPa,塑性应变迹和PEEQ单调性违例均为0。它的物理可接受性高于普通网络,但应力精度并未超过GRU,路径OOD误差仍为66.00 MPa。
50
+
51
+ ### 物理积分网络
52
+
53
+ 新增模型不再直接回归应力历程。网络预测塑性乘子初值,随后由可微的J2/Chaboche一致性方程进行局部校正;状态中显式保存无迹塑性应变、PEEQ和两组无迹背应力,应力由弹性关系重构。模型具有3,265个参数,并严格保证PEEQ单调、塑性不可压缩、非负塑性增量及关联流动结构。
54
+
55
+ 该模型在路径OOD上的应力RMSE为0.0237 MPa,相比最佳黑箱TCN的60.44 MPa降低三个数量级;121点降采样测试为1.44 MPa,仍明显优于黑箱网络。这一优势来自网络与数据生成器共享J2/Chaboche物理骨架,因此它应被视为“白盒物理融合上限/算法积分基准”,不能据此宣称任意物理约束网络都会优于数据驱动网络。下一阶段应在物理模型不完备、参数未知或实验数据条件下验证灰盒增益。
56
+
57
+ ### 结构有限元验证
58
+
59
+ 构建了12单元变截面杆,由原生AgentFEM Chaboche更新提供参考,在每个单元积分点嵌入神经本构,并通过全局平衡求解位移和反力。
60
+
61
+ - 轻缺口(截面削弱12%):GRU、物理状态GRU和物理积分网络均收敛;反力相对L2误差分别为25.5%、52.4%和1.65×10⁻⁷。
62
+ - 强缺口(截面削弱42%):GRU收敛但反力相对L2误差37.0%;物理状态GRU在第58步不收敛;物理积分网络反力相对L2误差2.39×10⁻⁷,并在完全卸载的第58步记录一次全局信赖域回退。
63
+
64
+ 结果同时保留成功与失败。它证明材料点RMSE无法替代结构级验证,也显示“输出端物理罚项”和“把物理积分写入架构”存在本质差别。严重工况仍需全局求解器回退,说明局部本构精度与全局求解稳健性是两项独立要求。
65
+
66
+ ## 三、理论叙事
67
+
68
+ ### 1. 本构记忆
69
+
70
+ 塑性状态不能写成当前应变的单值函数。点式MLP误差显著高于时序模型,为“历史变量不可省略”提供直接对照。
71
+
72
+ ### 2. OOD分层
73
+
74
+ - ID:相同路径族和参数域中的新轨迹;
75
+ - 参数OOD:路径形式已见,但初始屈服应力位于训练域两端;
76
+ - 路径OOD:完整路径族未进入训练,包含新的方向旋转与随机五维加载;
77
+ - 分辨率OOD:相同连续路径采用不同离散点数。
78
+
79
+ 把这些测试混为一个随机划分,会系统性高估工程可靠性。
80
+
81
+ ### 3. 物理约束的层级
82
+
83
+ 物理约束至少分为四层:状态合法性、热力学/屈服一致��、时间离散自洽性、结构求解器一致性。本工作由物理状态GRU覆盖第一层,由物理积分网络进一步覆盖前两层并显著改善第三层;结构门槛则表明第四层仍需要一致切线和全局稳健策略共同保障。
84
+
85
+ ### 4. 材料点到结构的误差放大
86
+
87
+ 结构平衡会将局部应力误差反馈到应变重分配;缺口越强,局部路径越容易偏离训练分布。切线误差还会影响Newton收敛。因此结构验证既评估预测精度,也评估算法切线、稳定性和适用域。
88
+
89
+ ## 四、文章定位
90
+
91
+ 当前成果已经足以形成一篇严谨的开放基准/计算力学研究论文,亮点是“严格区分ID与多类OOD、比较软约束与嵌入式物理、完成结构部署验证并报告求解器回退”。适合以基准、验证方法和工程可信性为核心投稿计算力学或工程AI期刊。
92
+
93
+ 若希望冲击CMAME等更高水平期刊,还需要把当前白盒物理积分网络推进为灰盒方法:故意采用不完备物理骨架,让网络学习未知演化项或潜在内部变量;加入实验或高保真微观数据;提供稳定且经数值差分核验的一致切线;并在二维或三维结构中验证几何、边界和路径OOD泛化。
94
+
95
+ ## 五、下一方法工作的最小闭环
96
+
97
+ 1. 将输入从总历程改为当前增量、旧状态和材料参数;
98
+ 2. 预测满足约束的状态增量,而不是任意时刻的绝对状态;
99
+ 3. 对同一连续路径随机改变步长,强制终态和中间态自洽;
100
+ 4. 用自动微分获得离散算法切线,并与固定旧状态数值差分比较;
101
+ 5. 在训练中加入一到两步全局平衡残差;
102
+ 6. 通过轻缺口、强缺口以及二维剪切/拉伸结构门槛;
103
+ 7. 报告精度、物理违例、收敛率、切线误差和端到端加速比。
104
+
105
+ 完成上述闭环后,文章贡献将从“高质量基准揭示问题”提升为“提出并验证解决方法”。
environment-reproduce.yml ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: agentfem-t2-reproduce
2
+ channels:
3
+ - conda-forge
4
+ dependencies:
5
+ - python=3.11
6
+ - fenics-dolfinx=0.11.0
7
+ - numpy>=2.0,<3
8
+ - scipy>=1.11,<2
9
+ - h5py>=3.10,<4
10
+ - matplotlib-base>=3.8,<4
11
+ - pytorch>=2.4
12
+ - pytest>=8,<10
13
+ - pip
14
+ - pip:
15
+ - git+https://github.com/haoming-luo/agentfem.git@058faecc05aeda143d014fd229401003a9258bbb
environment-use.yml ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ name: agentfem-t2-use
2
+ channels:
3
+ - conda-forge
4
+ dependencies:
5
+ - python=3.11
6
+ - numpy>=2.0,<3
7
+ - h5py>=3.10,<4
8
+ - matplotlib-base>=3.8,<4
9
+ - pytorch>=2.4
10
+ - pytest>=8,<10
models/t2_multiaxial_ood_v2/README.md ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # T2 multiaxial constitutive-memory model card
2
+
3
+ ## Intended use
4
+
5
+ These compact PyTorch models are reproducible research baselines for the
6
+ AgentFEM Multiaxial Constitutive Memory v2 dataset. They support comparisons of
7
+ history representation, loading-path OOD generalization, material-parameter
8
+ OOD generalization and temporal-resolution sensitivity.
9
+
10
+ They are not approved constitutive models for safety-critical analysis.
11
+
12
+ ## Files
13
+
14
+ Each `{protocol}_{model}.pt` checkpoint contains:
15
+
16
+ - architecture name and input size;
17
+ - weights (`state_dict`);
18
+ - train-only normalization statistics;
19
+ - protocol and split-field identity;
20
+ - material-parameter and Voigt ordering;
21
+ - random seed and parameter count.
22
+
23
+ The architecture definitions and training/evaluation code are in
24
+ `src/train_t2_multiaxial_models.py`. Training histories are stored as JSON.
25
+
26
+ ## Architectures
27
+
28
+ - `pointwise_mlp`: current strain and material parameters, no history.
29
+ - `gru`: two-layer causal GRU.
30
+ - `lstm`: two-layer causal LSTM.
31
+ - `causal_tcn`: four residual dilated causal-convolution blocks.
32
+ - `physics_state_gru`: predicts internal variables; enforces exactly
33
+ deviatoric plastic strain and nondecreasing PEEQ, and reconstructs stress
34
+ from elastic strain.
35
+ - `physics_integrator_nn`: predicts a bounded plastic-multiplier correction,
36
+ then applies differentiable J2/Chaboche consistency corrections. It stores
37
+ plastic strain, PEEQ and two backstress tensors explicitly.
38
+
39
+ ## Physics-integrator result
40
+
41
+ | Protocol | Stress RMSE | 121-state RMSE |
42
+ |---|---:|---:|
43
+ | ID | 0.0128 MPa | 1.08 MPa |
44
+ | Path OOD | 0.0237 MPa | 1.44 MPa |
45
+ | Parameter OOD | 0.0275 MPa | 1.14 MPa |
46
+
47
+ The new architecture is a white-box physics-fusion ceiling: its return-mapping
48
+ equations match the J2 and Chaboche families that generated the synthetic
49
+ data. The neural part learns a local integration correction rather than the
50
+ entire constitutive law. It is therefore a valid integration/deployment
51
+ benchmark, but not evidence that the same accuracy will transfer to unknown
52
+ materials or misspecified physics.
53
+
54
+ ## Critical limitations
55
+
56
+ - The best ID score is the GRU's 17.26 MPa RMSE; this does not transfer to
57
+ path OOD, where the best result is the TCN's 60.44 MPa.
58
+ - Sequence decimation substantially increases recurrent-model error.
59
+ - The physics-state GRU satisfies selected hard constraints but does not
60
+ provide a proven consistent algorithmic tangent.
61
+ - The physics-integrator NN uses a numerical discrete tangent. The severe
62
+ structural case required one recorded trust-region fallback at complete
63
+ unloading; its reaction relative-L2 error was 2.39e-7.
64
+ - Structural deployment amplified black-box errors; the severe-notch
65
+ physics-state GRU did not converge.
66
+ - Checkpoints use PyTorch serialization. Load only files obtained from this
67
+ repository and keep the architecture source beside the weights.
68
+
69
+ See `artifacts/t2_multiaxial_ood_v2/model_metrics.json` for all reported
70
+ metrics and the structural-validation directories for solver evidence.
models/t2_multiaxial_ood_v2/id_causal_tcn.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c078b5514c8fc2245be0dcb5a7051a9e229f4168dc4ace5753ef8bbc6dd775d4
3
+ size 267239
models/t2_multiaxial_ood_v2/id_causal_tcn_history.json ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "epoch": 1,
4
+ "train_loss": 0.520246310159564,
5
+ "validation_loss": 0.23935295641422272
6
+ },
7
+ {
8
+ "epoch": 2,
9
+ "train_loss": 0.16315131541341543,
10
+ "validation_loss": 0.15028159320354462
11
+ },
12
+ {
13
+ "epoch": 3,
14
+ "train_loss": 0.1111618389065067,
15
+ "validation_loss": 0.11974978446960449
16
+ },
17
+ {
18
+ "epoch": 4,
19
+ "train_loss": 0.09225699336578448,
20
+ "validation_loss": 0.10635652393102646
21
+ },
22
+ {
23
+ "epoch": 5,
24
+ "train_loss": 0.08524151394764583,
25
+ "validation_loss": 0.10028232634067535
26
+ },
27
+ {
28
+ "epoch": 6,
29
+ "train_loss": 0.07768653038268288,
30
+ "validation_loss": 0.09240853786468506
31
+ },
32
+ {
33
+ "epoch": 7,
34
+ "train_loss": 0.07283600233495235,
35
+ "validation_loss": 0.08987530320882797
36
+ },
37
+ {
38
+ "epoch": 8,
39
+ "train_loss": 0.071241222632428,
40
+ "validation_loss": 0.08638766407966614
41
+ },
42
+ {
43
+ "epoch": 9,
44
+ "train_loss": 0.06595686404034495,
45
+ "validation_loss": 0.08314426988363266
46
+ },
47
+ {
48
+ "epoch": 10,
49
+ "train_loss": 0.06490953639149666,
50
+ "validation_loss": 0.08100694417953491
51
+ },
52
+ {
53
+ "epoch": 11,
54
+ "train_loss": 0.06278126128017902,
55
+ "validation_loss": 0.08084318786859512
56
+ },
57
+ {
58
+ "epoch": 12,
59
+ "train_loss": 0.05896675509090225,
60
+ "validation_loss": 0.07517626881599426
61
+ },
62
+ {
63
+ "epoch": 13,
64
+ "train_loss": 0.055060570438702904,
65
+ "validation_loss": 0.07204269617795944
66
+ },
67
+ {
68
+ "epoch": 14,
69
+ "train_loss": 0.0550352114563187,
70
+ "validation_loss": 0.07488186657428741
71
+ },
72
+ {
73
+ "epoch": 15,
74
+ "train_loss": 0.052463256288319826,
75
+ "validation_loss": 0.06767647713422775
76
+ },
77
+ {
78
+ "epoch": 16,
79
+ "train_loss": 0.0507348362977306,
80
+ "validation_loss": 0.07125037908554077
81
+ },
82
+ {
83
+ "epoch": 17,
84
+ "train_loss": 0.05015234788879752,
85
+ "validation_loss": 0.06936879456043243
86
+ },
87
+ {
88
+ "epoch": 18,
89
+ "train_loss": 0.048359594540670514,
90
+ "validation_loss": 0.06636278331279755
91
+ },
92
+ {
93
+ "epoch": 19,
94
+ "train_loss": 0.04642130403469006,
95
+ "validation_loss": 0.06443136930465698
96
+ },
97
+ {
98
+ "epoch": 20,
99
+ "train_loss": 0.04528034416337808,
100
+ "validation_loss": 0.07277429103851318
101
+ },
102
+ {
103
+ "epoch": 21,
104
+ "train_loss": 0.04685052034134666,
105
+ "validation_loss": 0.06481491774320602
106
+ },
107
+ {
108
+ "epoch": 22,
109
+ "train_loss": 0.04537177776607374,
110
+ "validation_loss": 0.062488581985235214
111
+ },
112
+ {
113
+ "epoch": 23,
114
+ "train_loss": 0.040842023522903524,
115
+ "validation_loss": 0.0594027154147625
116
+ },
117
+ {
118
+ "epoch": 24,
119
+ "train_loss": 0.039278608125944935,
120
+ "validation_loss": 0.05833548679947853
121
+ },
122
+ {
123
+ "epoch": 25,
124
+ "train_loss": 0.04030861690019568,
125
+ "validation_loss": 0.059121910482645035
126
+ },
127
+ {
128
+ "epoch": 26,
129
+ "train_loss": 0.03749564460789164,
130
+ "validation_loss": 0.05871205776929855
131
+ },
132
+ {
133
+ "epoch": 27,
134
+ "train_loss": 0.03777999981927375,
135
+ "validation_loss": 0.057577554136514664
136
+ },
137
+ {
138
+ "epoch": 28,
139
+ "train_loss": 0.038198326636726655,
140
+ "validation_loss": 0.05568905174732208
141
+ },
142
+ {
143
+ "epoch": 29,
144
+ "train_loss": 0.03582317933129767,
145
+ "validation_loss": 0.05356375128030777
146
+ },
147
+ {
148
+ "epoch": 30,
149
+ "train_loss": 0.035069097221518554,
150
+ "validation_loss": 0.059687290340662
151
+ },
152
+ {
153
+ "epoch": 31,
154
+ "train_loss": 0.03609969102156659,
155
+ "validation_loss": 0.052147723734378815
156
+ },
157
+ {
158
+ "epoch": 32,
159
+ "train_loss": 0.03294815899183353,
160
+ "validation_loss": 0.052073314785957336
161
+ },
162
+ {
163
+ "epoch": 33,
164
+ "train_loss": 0.032148236564050116,
165
+ "validation_loss": 0.053652044385671616
166
+ },
167
+ {
168
+ "epoch": 34,
169
+ "train_loss": 0.03145686265391608,
170
+ "validation_loss": 0.05307731777429581
171
+ },
172
+ {
173
+ "epoch": 35,
174
+ "train_loss": 0.030813092753912013,
175
+ "validation_loss": 0.05330563336610794
176
+ }
177
+ ]
models/t2_multiaxial_ood_v2/id_gru.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b346b3f373d5bea97febc396cb0f4f4c8db59041c97dcea345e975624605d0fb
3
+ size 212293
models/t2_multiaxial_ood_v2/id_gru_history.json ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "epoch": 1,
4
+ "train_loss": 0.8190490131576856,
5
+ "validation_loss": 0.7432200908660889
6
+ },
7
+ {
8
+ "epoch": 2,
9
+ "train_loss": 0.6653396102289358,
10
+ "validation_loss": 0.6271824836730957
11
+ },
12
+ {
13
+ "epoch": 3,
14
+ "train_loss": 0.523029014468193,
15
+ "validation_loss": 0.4923483729362488
16
+ },
17
+ {
18
+ "epoch": 4,
19
+ "train_loss": 0.4074751523633798,
20
+ "validation_loss": 0.3707224726676941
21
+ },
22
+ {
23
+ "epoch": 5,
24
+ "train_loss": 0.26617325966556865,
25
+ "validation_loss": 0.24201348423957825
26
+ },
27
+ {
28
+ "epoch": 6,
29
+ "train_loss": 0.1702969092875719,
30
+ "validation_loss": 0.16556040942668915
31
+ },
32
+ {
33
+ "epoch": 7,
34
+ "train_loss": 0.12546609776715437,
35
+ "validation_loss": 0.13362598419189453
36
+ },
37
+ {
38
+ "epoch": 8,
39
+ "train_loss": 0.09630454859385888,
40
+ "validation_loss": 0.09932821989059448
41
+ },
42
+ {
43
+ "epoch": 9,
44
+ "train_loss": 0.07659551035612822,
45
+ "validation_loss": 0.08940248936414719
46
+ },
47
+ {
48
+ "epoch": 10,
49
+ "train_loss": 0.06631685762355725,
50
+ "validation_loss": 0.0795120894908905
51
+ },
52
+ {
53
+ "epoch": 11,
54
+ "train_loss": 0.06089257759352525,
55
+ "validation_loss": 0.07226704806089401
56
+ },
57
+ {
58
+ "epoch": 12,
59
+ "train_loss": 0.0542294104428341,
60
+ "validation_loss": 0.06714063137769699
61
+ },
62
+ {
63
+ "epoch": 13,
64
+ "train_loss": 0.05126884990992645,
65
+ "validation_loss": 0.06375377625226974
66
+ },
67
+ {
68
+ "epoch": 14,
69
+ "train_loss": 0.04738584657510122,
70
+ "validation_loss": 0.06645862013101578
71
+ },
72
+ {
73
+ "epoch": 15,
74
+ "train_loss": 0.0468538097726802,
75
+ "validation_loss": 0.059026267379522324
76
+ },
77
+ {
78
+ "epoch": 16,
79
+ "train_loss": 0.042510446005811296,
80
+ "validation_loss": 0.05520983412861824
81
+ },
82
+ {
83
+ "epoch": 17,
84
+ "train_loss": 0.04032765685891112,
85
+ "validation_loss": 0.053845543414354324
86
+ },
87
+ {
88
+ "epoch": 18,
89
+ "train_loss": 0.03821880021132529,
90
+ "validation_loss": 0.0499015748500824
91
+ },
92
+ {
93
+ "epoch": 19,
94
+ "train_loss": 0.03617646490844587,
95
+ "validation_loss": 0.05045943334698677
96
+ },
97
+ {
98
+ "epoch": 20,
99
+ "train_loss": 0.03531732417953511,
100
+ "validation_loss": 0.04865029454231262
101
+ },
102
+ {
103
+ "epoch": 21,
104
+ "train_loss": 0.03382568100156883,
105
+ "validation_loss": 0.049336206167936325
106
+ },
107
+ {
108
+ "epoch": 22,
109
+ "train_loss": 0.0339934304356575,
110
+ "validation_loss": 0.04705159366130829
111
+ },
112
+ {
113
+ "epoch": 23,
114
+ "train_loss": 0.03202988890310129,
115
+ "validation_loss": 0.046920355409383774
116
+ },
117
+ {
118
+ "epoch": 24,
119
+ "train_loss": 0.030802785651758313,
120
+ "validation_loss": 0.04801936075091362
121
+ },
122
+ {
123
+ "epoch": 25,
124
+ "train_loss": 0.030567849054932594,
125
+ "validation_loss": 0.04281919449567795
126
+ },
127
+ {
128
+ "epoch": 26,
129
+ "train_loss": 0.030341880784059565,
130
+ "validation_loss": 0.04389240965247154
131
+ },
132
+ {
133
+ "epoch": 27,
134
+ "train_loss": 0.02857093675993383,
135
+ "validation_loss": 0.041295621544122696
136
+ },
137
+ {
138
+ "epoch": 28,
139
+ "train_loss": 0.026360594977935154,
140
+ "validation_loss": 0.04069531336426735
141
+ },
142
+ {
143
+ "epoch": 29,
144
+ "train_loss": 0.026334694509084027,
145
+ "validation_loss": 0.0402633473277092
146
+ },
147
+ {
148
+ "epoch": 30,
149
+ "train_loss": 0.025304252825056512,
150
+ "validation_loss": 0.03932539373636246
151
+ },
152
+ {
153
+ "epoch": 31,
154
+ "train_loss": 0.025048449402675033,
155
+ "validation_loss": 0.037513844668865204
156
+ },
157
+ {
158
+ "epoch": 32,
159
+ "train_loss": 0.022826850647106767,
160
+ "validation_loss": 0.03704925626516342
161
+ },
162
+ {
163
+ "epoch": 33,
164
+ "train_loss": 0.02272752233936141,
165
+ "validation_loss": 0.035603757947683334
166
+ },
167
+ {
168
+ "epoch": 34,
169
+ "train_loss": 0.02206879233320554,
170
+ "validation_loss": 0.03528529033064842
171
+ },
172
+ {
173
+ "epoch": 35,
174
+ "train_loss": 0.021932116088767845,
175
+ "validation_loss": 0.03687312826514244
176
+ }
177
+ ]
models/t2_multiaxial_ood_v2/id_lstm.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c058e02c5283f9367b97c68bcdcf66d9319462a2d60b23e7587162a09cd1da5b
3
+ size 280709
models/t2_multiaxial_ood_v2/id_lstm_history.json ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "epoch": 1,
4
+ "train_loss": 0.8774643167853355,
5
+ "validation_loss": 0.7828194499015808
6
+ },
7
+ {
8
+ "epoch": 2,
9
+ "train_loss": 0.685258706410726,
10
+ "validation_loss": 0.6396182775497437
11
+ },
12
+ {
13
+ "epoch": 3,
14
+ "train_loss": 0.5001022045811018,
15
+ "validation_loss": 0.4530872702598572
16
+ },
17
+ {
18
+ "epoch": 4,
19
+ "train_loss": 0.35807362881799537,
20
+ "validation_loss": 0.34114503860473633
21
+ },
22
+ {
23
+ "epoch": 5,
24
+ "train_loss": 0.2592892286678155,
25
+ "validation_loss": 0.2634303867816925
26
+ },
27
+ {
28
+ "epoch": 6,
29
+ "train_loss": 0.192195700481534,
30
+ "validation_loss": 0.2238457202911377
31
+ },
32
+ {
33
+ "epoch": 7,
34
+ "train_loss": 0.15395612021287283,
35
+ "validation_loss": 0.18534839153289795
36
+ },
37
+ {
38
+ "epoch": 8,
39
+ "train_loss": 0.1320283099388083,
40
+ "validation_loss": 0.1692781001329422
41
+ },
42
+ {
43
+ "epoch": 9,
44
+ "train_loss": 0.11203321519618233,
45
+ "validation_loss": 0.14782389998435974
46
+ },
47
+ {
48
+ "epoch": 10,
49
+ "train_loss": 0.09273092200358708,
50
+ "validation_loss": 0.13225607573986053
51
+ },
52
+ {
53
+ "epoch": 11,
54
+ "train_loss": 0.08526860006774466,
55
+ "validation_loss": 0.13516588509082794
56
+ },
57
+ {
58
+ "epoch": 12,
59
+ "train_loss": 0.08009990180532138,
60
+ "validation_loss": 0.1203681007027626
61
+ },
62
+ {
63
+ "epoch": 13,
64
+ "train_loss": 0.07252205737556021,
65
+ "validation_loss": 0.11525837332010269
66
+ },
67
+ {
68
+ "epoch": 14,
69
+ "train_loss": 0.06653289139891665,
70
+ "validation_loss": 0.10819903761148453
71
+ },
72
+ {
73
+ "epoch": 15,
74
+ "train_loss": 0.05912485811859369,
75
+ "validation_loss": 0.10831790417432785
76
+ },
77
+ {
78
+ "epoch": 16,
79
+ "train_loss": 0.0566910100945582,
80
+ "validation_loss": 0.10237377136945724
81
+ },
82
+ {
83
+ "epoch": 17,
84
+ "train_loss": 0.05357912927865982,
85
+ "validation_loss": 0.09337074309587479
86
+ },
87
+ {
88
+ "epoch": 18,
89
+ "train_loss": 0.04630529647693038,
90
+ "validation_loss": 0.0894143283367157
91
+ },
92
+ {
93
+ "epoch": 19,
94
+ "train_loss": 0.044135728695740305,
95
+ "validation_loss": 0.08880313485860825
96
+ },
97
+ {
98
+ "epoch": 20,
99
+ "train_loss": 0.04071127778540055,
100
+ "validation_loss": 0.0864553451538086
101
+ },
102
+ {
103
+ "epoch": 21,
104
+ "train_loss": 0.03878524092336496,
105
+ "validation_loss": 0.08670862019062042
106
+ },
107
+ {
108
+ "epoch": 22,
109
+ "train_loss": 0.038949551138406,
110
+ "validation_loss": 0.08375247567892075
111
+ },
112
+ {
113
+ "epoch": 23,
114
+ "train_loss": 0.03704802032249669,
115
+ "validation_loss": 0.0790134146809578
116
+ },
117
+ {
118
+ "epoch": 24,
119
+ "train_loss": 0.03656226877743999,
120
+ "validation_loss": 0.08138751238584518
121
+ },
122
+ {
123
+ "epoch": 25,
124
+ "train_loss": 0.03616096063827475,
125
+ "validation_loss": 0.08085963875055313
126
+ },
127
+ {
128
+ "epoch": 26,
129
+ "train_loss": 0.03707934178722402,
130
+ "validation_loss": 0.07809192687273026
131
+ },
132
+ {
133
+ "epoch": 27,
134
+ "train_loss": 0.03426934087959429,
135
+ "validation_loss": 0.07316715270280838
136
+ },
137
+ {
138
+ "epoch": 28,
139
+ "train_loss": 0.031773798555756606,
140
+ "validation_loss": 0.0726182609796524
141
+ },
142
+ {
143
+ "epoch": 29,
144
+ "train_loss": 0.03107247066994508,
145
+ "validation_loss": 0.07648567110300064
146
+ },
147
+ {
148
+ "epoch": 30,
149
+ "train_loss": 0.030413803023596604,
150
+ "validation_loss": 0.0704488530755043
151
+ },
152
+ {
153
+ "epoch": 31,
154
+ "train_loss": 0.026944131978477042,
155
+ "validation_loss": 0.06747745722532272
156
+ },
157
+ {
158
+ "epoch": 32,
159
+ "train_loss": 0.025804441887885332,
160
+ "validation_loss": 0.06794652342796326
161
+ },
162
+ {
163
+ "epoch": 33,
164
+ "train_loss": 0.025182920585696895,
165
+ "validation_loss": 0.06860146671533585
166
+ },
167
+ {
168
+ "epoch": 34,
169
+ "train_loss": 0.024644560801486175,
170
+ "validation_loss": 0.0657595545053482
171
+ },
172
+ {
173
+ "epoch": 35,
174
+ "train_loss": 0.024648283375427127,
175
+ "validation_loss": 0.0651465654373169
176
+ }
177
+ ]
models/t2_multiaxial_ood_v2/id_normalization.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "normalization": {
3
+ "backstress_scale": [
4
+ 332514720.0,
5
+ 289896480.0,
6
+ 271324704.0,
7
+ 275844000.0,
8
+ 275541760.0,
9
+ 275399040.0
10
+ ],
11
+ "parameter_mean": [
12
+ 189718740992.0,
13
+ 0.2956160604953766,
14
+ 313079200.0,
15
+ 2132017792.0,
16
+ 18812704768.0,
17
+ 34.88433837890625,
18
+ 4281594624.0,
19
+ 5.357907772064209,
20
+ 35515076.0,
21
+ 5.472315311431885
22
+ ],
23
+ "parameter_std": [
24
+ 23155456000.0,
25
+ 0.026136940345168114,
26
+ 77052360.0,
27
+ 2609331712.0,
28
+ 20978900992.0,
29
+ 40.323970794677734,
30
+ 5035721216.0,
31
+ 6.5961689949035645,
32
+ 40976276.0,
33
+ 6.580267906188965
34
+ ],
35
+ "peeq_scale": [
36
+ 0.09643255919218063
37
+ ],
38
+ "plastic_scale": [
39
+ 0.01167806051671505,
40
+ 0.011851734481751919,
41
+ 0.011857790872454643,
42
+ 0.009260914288461208,
43
+ 0.008840655907988548,
44
+ 0.008795959874987602
45
+ ],
46
+ "strain_mean": [
47
+ 0.0006159174372442067,
48
+ 0.00018006325990427285,
49
+ -0.0007959806825965643,
50
+ 0.0005156935076229274,
51
+ 0.00010772027599159628,
52
+ 5.708914977731183e-05
53
+ ],
54
+ "strain_std": [
55
+ 0.0038143135607242584,
56
+ 0.0026040070224553347,
57
+ 0.0024796186480671167,
58
+ 0.0025680263061076403,
59
+ 0.0009052220266312361,
60
+ 0.000912147224880755
61
+ ],
62
+ "stress_mean": [
63
+ 3622040.25,
64
+ 1752063.625,
65
+ -5374104.5,
66
+ 4515719.0,
67
+ 511102.1875,
68
+ 1951768.625
69
+ ],
70
+ "stress_std": [
71
+ 171135232.0,
72
+ 109217288.0,
73
+ 113089080.0,
74
+ 129707432.0,
75
+ 50405604.0,
76
+ 48862108.0
77
+ ]
78
+ },
79
+ "parameter_names": [
80
+ "young_pa",
81
+ "poisson",
82
+ "yield_stress_pa",
83
+ "hardening_modulus_pa",
84
+ "backstress_c1_pa",
85
+ "backstress_gamma1",
86
+ "backstress_c2_pa",
87
+ "backstress_gamma2",
88
+ "isotropic_saturation_pa",
89
+ "isotropic_rate"
90
+ ],
91
+ "protocol": "id",
92
+ "split_field": "split_id"
93
+ }
models/t2_multiaxial_ood_v2/id_physics_integrator_nn.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e46c537a2963fbe6bea4c215f4118ca0b496ebfda8f5e2c0cf31a5ba455b0430
3
+ size 16781