Convert per-case Abaqus FEA outputs into ML-ready (X, Y) wide-table CSVs. Pivots irregular FEA mesh node displacements onto a regular N×N grid via direct binning (structured mesh) or bilinear resampling, picks the final frame as the deformation target, and aggregates across many cases into X_amplitude.csv (design vectors) + Y_grid_uz.csv (flattened grid displacement). Use when the user has a folder of completed FEA cases and wants to train a Ridge / MLP / Gaussian Process surrogate on the (input → displacement field) mapping.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Convert per-case Abaqus FEA outputs into ML-ready (X, Y) wide-table CSVs. Pivots irregular FEA mesh node displacements onto a regular N×N grid via direct binning (structured mesh) or bilinear resampling, picks the final frame as the deformation target, and aggregates across many cases into X_amplitude.csv (design vectors) + Y_grid_uz.csv (flattened grid displacement). Use when the user has a folder of completed FEA cases and wants to train a Ridge / MLP / Gaussian Process surrogate on the (input → displacement field) mapping.
The post-processing half of a surrogate-model pipeline. Takes a folder of completed Abaqus FEA cases (typically produced by the abaqus-lhs-batch-dataset skill) and produces two flat CSVs ready for numpy.loadtxt / pandas.read_csv / scikit-learn:
Time-series outputs where you need every frame (this skill takes the final frame by default)
Stress / strain field extraction (this skill is targeted at displacement; extending to other fields is straightforward, see below)
The Two Mesh-to-Grid Paths
Path A — Structured FEA mesh (preferred, fast)
The FEA mesh is a regular grid (e.g. 81×81 nodes for a square membrane), and you want to either keep it or downsample to a smaller learning grid. This is the case for membrane / plate problems where you set the mesh seed to give exact node spacing.
Steps:
Detect mesh size: unique(x0) and unique(y0) from node_displacement.csv give MESH_N
Use rounded (x0, y0) as a hash key → directly bin each node into a (MESH_N, MESH_N) array
If target_N != MESH_N: bilinear resample to (target_N, target_N) using precomputed weights (10-100× faster than )
scipy.interpolate
Flatten row-major to length target_N²
Latency: ~50 ms / case. 1000 cases = ~1 minute.
Path B — Unstructured FEA mesh (fallback)
The FEA mesh is unstructured (tet / triangular / mixed). Need scattered interpolation.
Reject the run if < 80% of planned cases completed — surface the failure rate to the user.
Step 2 — Detect mesh structure (one-time)
Open the first completed case's node_displacement.csv. Count unique x0 and y0 (rounded to 2 decimals to absorb floating-point noise). If len(unique_x) * len(unique_y) == NODES_PER_FRAME, the mesh is structured (Path A). Otherwise unstructured (Path B).
Step 3 — Per-case extraction
For each completed case (in parallel via ThreadPoolExecutor or serially):
Read input vector → shape (D,)
Read final frame from node_displacement.csv:
Fast path: tail -NODES_PER_FRAME (Linux/macOS) or read whole file + filter by frame_id == max_frame_id
Extract (x0, y0, uz) columns
Apply Path A or Path B to produce (target_N, target_N) displacement field
Flatten row-major to (target_N²,)
Append to in-memory accumulator
Step 4 — Aggregate + deduplicate
Stack all input vectors into X of shape (N_completed, D)
Stack all flattened grids into Y of shape (N_completed, target_N²)
Optional: deduplicate identical input vectors using numpy.unique with return_index=True. Why dedupe: identical inputs but slightly different outputs (numerical noise) can confuse the regularization; pick one representative per input.
Y percentiles: p1, p50, p99 — sanity check the displacement range is reasonable (no nan / inf)
Top-5 largest |uz| cases — visually inspect to make sure they aren't obviously diverged
Critical Implementation Details
1. Node coordinates may not align exactly across cases
Floating-point: a node nominally at x = 1.0 may show up as 0.99999998 in one case and 1.00000002 in another. Always roundx0, y0 to the nearest mesh_pitch / 100 (e.g. np.round(x0, 2) for mm-scale meshes). Otherwise the unique-coordinate detection will fail.
2. The final frame is not always the last row
Some Abaqus output configurations write the initial (zero) frame at the start, the loading frames, and a final equilibrium frame. Always select by max(frame_id), not by tail-N.
3. Bilinear resample weights are precomputable
For a fixed MESH_N → target_N mapping, the per-target-cell (idx0, idx1, frac) weights only depend on the grid sizes. Precompute once and reuse for all cases:
Then per-case bilinear is a vectorized numpy operation, no Python loop over grid cells.
4. Handle large node_displacement.csv files efficiently
Each file is per-frame × per-node × 12 columns. For 81×81 mesh × 20 frames × 12 cols × ~30 bytes ≈ 47 MB per case. Reading the full file with csv.DictReader is slow.
Fast read: skip to the last NODES_PER_FRAME lines via subprocess.run(["tail", "-N", path]) (Linux/macOS only). On Windows, use numpy.genfromtxt with skip_header= set to (num_frames - 1) * NODES_PER_FRAME + 1. Even faster: have the FEA solver script output ONLY the final frame to a separate file like final_frame.csv, eliminating the parse step.
5. Memory budget
Y for 1000 samples × 21×21 grid = 1000 × 441 × 8 bytes = 3.5 MB → fine in memory
Y for 10,000 samples × 81×81 grid = 10000 × 6561 × 8 bytes = 525 MB → still in memory, but write CSV incrementally
For larger sizes: write each row to disk immediately (no in-memory accumulator)
6. Failure modes
Mismatched dimensions: a case has len(unique_x) != MESH_N — log and skip, don't crash
NaN in uz: solver diverged silently. Should never happen if dataset_index.csv says "completed", but check np.isfinite(uz).all() per case anyway
Missing input_vector.csv: fall back to parsing ForceAmplitude.dat directly (the *Amplitude value lines)
Reference Implementation
references/extract_grid.py (~250 lines) — full pipeline as a CLI:
After producing X_amplitude.csv and Y_grid_uz.csv:
Shape match: len(X) == len(Y) and both index by the same sample_id
Y range: Y.abs().max() between sane bounds (not inf, not 0)
Linear fit baseline: train a Ridge with alpha=1.0 and check R² on a 80/20 split. Should be > 0.9 for well-behaved problems. R² < 0.5 implies the design space is too noisy or the FEA setup is suspect — surface this to the user.
Dedup ratio: if n_dedup / n_total < 0.7, the design space sampler is producing too many duplicates — narrow the precision threshold or use a different sampler.
