Generate an Abaqus FEA training dataset for surrogate / ML models. Latin Hypercube Sampling (or sparse-pattern sampling) over a parameterized design vector, one case folder per sample, batch-submit Abaqus jobs via subprocess, recover from crashes, and write a unified dataset index. Use when the user wants to "build a training set for a surrogate model", "sweep design parameters in Abaqus", "run N FEA simulations", or "sample a design space".
Instrucciones de origen · Vista previa de solo lectura
name
abaqus-lhs-batch-dataset
description
Generate an Abaqus FEA training dataset for surrogate / ML models. Latin Hypercube Sampling (or sparse-pattern sampling) over a parameterized design vector, one case folder per sample, batch-submit Abaqus jobs via subprocess, recover from crashes, and write a unified dataset index. Use when the user wants to "build a training set for a surrogate model", "sweep design parameters in Abaqus", "run N FEA simulations", or "sample a design space".
End-to-end workflow for producing FEA training datasets from a parameterized Abaqus model. Designed for surrogate model training (Ridge, MLP, Gaussian Process, etc.) where you need hundreds to thousands of FEA samples covering a design space.
When to Use This Skill
Activate when the user wants to:
Build a training set for a surrogate / ML model from Abaqus FEA
Sweep a multi-dimensional design vector (e.g. force amplitudes, geometry parameters, material properties)
Run N independent Abaqus jobs in batch with progress tracking and crash recovery
Generate inputs for inverse-design / topology-optimization pipelines
Do NOT use this skill for:
Single-case Abaqus runs (use abaqus-job instead)
Parametric sweeps that don't need ML-ready CSV output (use abaqus-job + manual analysis)
Optimization loops where each evaluation depends on the previous one (use a different driver)
The pattern is template-driven: you keep one canonical case with all the static input files (geometry, material, mesh, step parameters), and per sample you only overwrite the design vector file (e.g. ForceAmplitude.dat).
Required Inputs
The user must provide:
template_dir/ — A folder containing one working Abaqus case with all the static input files. Verify by running it once standalone and confirming the .odb is produced. Required files vary by setup, but typically include:
solver_script.py — The Abaqus/Python solver script that:
Reads the template .dat files
Reads the per-case design file (e.g. ForceAmplitude.dat)
Builds the model, submits the job, opens the ODB, and writes a per-node CSV (e.g. node_displacement.csv with columns frame_id, time, node_id, x0, y0, z0, ux, uy, uz, x_def, y_def, z_def)
Is invoked via abaqus cae noGUI="<path>" from inside the case dir
Design dimension D — number of free parameters per sample (e.g. 16 channels)
Sampling strategy + bounds (e.g. LHS in [-0.5, +0.5]^16)
Sampling Strategies
Pick one (or combine) based on the design space:
Strategy
When to use
Implementation
LHS (Latin Hypercube)
Dense, isotropic coverage of a continuous box
Stratified per-dim shuffle (no scipy needed) — see references/sampling.py
Uniform random
Baseline / sanity check
random.uniform(lo, hi) per dim
Sparse-k
Inverse problems with sparsity prior — k of D dims are non-zero
Random k-subset selection with discrete amplitude levels
Build a list of (design_id, vector) tuples per the chosen strategy
Deduplicate (use a string key from the rounded values to drop near-duplicates)
Write sample_design.csv with header design_id, active_count, amplitude_1, ..., amplitude_D
Step 3 — Per-sample case execution
For each sample i = 1..N:
case_dir = dataset_root / f"sample_{i:05d}"
Copy all files from template_dir/ into case_dir/ (shutil.copy2)
OverwriteForceAmplitude.dat in case_dir/ with the new design vector — this is the only file that changes between cases. Format (Abaqus *Amplitude blocks):
Before running a case, check if dataset_index.csv already has a completed row for that sample_id → skip if --reuse-existing
For failed rows, optionally retry with longer timeout
Step 5 — Report
Print summary: N completed / M total, mean elapsed = X s
List failed sample IDs for manual inspection of run_stderr.log
Critical Implementation Details
1. Working directory MUST be the case dir
Abaqus reads *.dat files relative to the launch directory. Always set cwd=case_dir in subprocess.run. NEVER invoke the solver from the dataset root or template dir.
2. Abaqus must be on PATH
Verify with abaqus --help before starting. On Windows, the install adds C:\SIMULIA\Commands to PATH.
3. CPU contention
Each abaqus cae job uses 1 CPU by default. Running batch jobs serially is the simplest correct approach. For parallelism, ensure numCpus * concurrent_jobs ≤ physical_cores and watch out for license token contention.
4. Disk usage
Each case generates ~50-200 MB (.odb + .dat + .csv + lockfiles). For 1000 samples, plan for 100-200 GB. Optionally delete .odb and intermediate files after extracting the per-node CSV (see the abaqus-odb-to-grid-csv skill).
5. Failure modes (in order of frequency)
License timeout: solver hangs waiting for a token → use Abaqus's lic_check_freq env var or detect via *** ERROR: TIMED OUT WAITING FOR LICENSE in stderr
Convergence failure: nonlinear solve diverges (typical with extreme amplitudes) → status = "failed", check .msg file
Mesh distortion: too-large displacements warp elements past Jacobian threshold → reduce amplitude bounds or refine mesh
File lock / leftover .lck files: previous crashed run left *.lck → script should rm -f *.lck before submitting
Time budget: a single case taking > timeout_s → mark as timeout, increase per-case budget for retry
6. Reproducibility
Always seed the RNG (random.Random(seed)) and write the seed to dataset_meta.json. LHS with the same seed + same N gives identical samples.
Reference Implementation
A complete, dependency-free Python implementation is in references/batch_runner.py (~400 lines). It's parameterized via argparse so it works as a CLI:
The next step in the pipeline is typically the abaqus-odb-to-grid-csv skill, which collapses each node_displacement.csv into a single row of a wide-table Y_grid_uz.csv (sample_id × N²), giving you ML-ready (X, Y) matrices.
Quick Sanity Checks
After a run completes, ask the agent to verify:
Index integrity: pandas.read_csv("dataset_index.csv") — count status == "completed"
Sample diversity: max-pairwise-correlation of sample_design.csv rows — should be < 0.7 for LHS
Output range: dataset_index.csv["max_abs_uz"] distribution — should span at least 1 order of magnitude (else the design space is too narrow)
Failed cases: cat sample_XXXXX/run_stderr.log for the first 3 failures
If a high failure rate (> 5%) appears, narrow the design bounds or check the solver script — don't blindly retry.
