FlowMLLab v1.10.0 - Transformer and language-modeling laboratories

FlowMLLab v1.10.0

This release adds Weeks 17–22 to the established CFD/SciML course and publishes
the corrections following independent review. The course now has 50 notebooks
and 33 lecture PDFs, with all modules linked from the main course table.

• Six CPU laboratories cover Q/K/V attention, next-character language modeling,
  tokenization and sensor design, missing-observation reconstruction, autonomous
  prediction and transfer/information-budget audits.
• Each new week includes four coding exercises, an executed student notebook,
  separate instructor solutions, an editable lecture and a five-page PDF.
• The numerical comparisons include classical and neural baselines, saved
  representation diagnostics, fully counted target labels, matched-update
  controls and a paired three-seed sensor-augmentation ablation. Checkpoint
  verification reports strict and declared scientific criteria separately.
• Six homepage/weekly-page previews now use native-grid CFD velocity/vorticity
  contours, sensor markers and absolute spatial error. Shared color scales,
  fixed snapshots and a provenance manifest support direct interpretation.
• Python distribution, runtime, citation and Zenodo metadata consistently report
  version 1.10.0. The portable Weeks 17–22 package includes the new landing pages
  and contour assets, together with its data, evidence and editable sources.

Existing prediction/checkpoint archives and original instructional figures are
preserved. The source release and portable package retain their checksums and
reproduction instructions; older large CFD inputs remain available at their
original, explicitly pinned releases.

Validation covers the package/release gates, scientific evidence, 24 coding
exercises and semantic mutation checks, independent checkpoint reconstruction,
the paired augmentation evidence, and all 12 new notebooks in fresh kernels.
See docs/TRANSFORMER_SECOND_REVIEW_RESPONSE.md and the GitHub Actions runs for
the released revision. The contour update separately checked 20 numerical
reference/prediction panels, displayed snapshot norms and 383 local links.

Interactive Colab sessions and a representative student workload/grading pilot
remain open classroom-acceptance checks. Previously inspected coarse LBM wakes
are retained evaluation cases, not new blind tests or grid-independence
evidence. The character language model and continuous-state CFD decoder remain
separate experiments; matched optimizer-update ceilings do not establish equal
FLOPs or source information. The earlier NASA reproduction limitations remain
unchanged. This is an open teaching/software release, not institutional course
approval or a claim of new scientific superiority.

FlowMLLab v1.9.1 - Consistent course homepage

FlowMLLab v1.9.1

The homepage now presents Weeks 15 and 16 in the same introduction, problem, CFD/data, learning-method and results format used by the preceding modules. Reuse, citation and license information follow Week 16 at the end of the page.

Week 16 now links the checked SU2 8.0.1 fields and reports the current neural errors and direct-CFD design gains. Stale SU2 8.5.0 headline numbers are removed. NASA remains a failed, deferred research extension. No CFD, training data, weights or notebook results were changed in this documentation release.

FlowMLLab v1.9.0 - Week 16 validated educational core

FlowMLLab v1.9.0 — Week 16 reference validation

This release covers the physically checked teaching-body CFD, neural surrogate, retained design comparisons and Taylor–Maccoll verification. NASA validation is explicitly deferred and is not part of the accepted results.

Original NASA SEEB-ALR geometry, experimental records and LAVA reference remain available for study. The 8,000-iteration SU2 8.0.1 replay did not converge (density residual log10 −2.3811) and failed physical checks (41.24% stagnation-pressure error, 25.48% maximum total-enthalpy deviation). No NASA CFD agreement is claimed. Further NASA work is a separate research task.

The notebook independently integrates the Taylor-Maccoll cone equations and distinguishes experimental CFD validation from neural testing. The new SU2 8.0.1 cone study has 1.1616% finest-grid pressure error and 1.0064% last-two pressure change. Ten fresh design checks give 20.6646% peak-pressure reduction and 3.5218% pressure-drag reduction on the finer mesh.

All 44 original training/validation/test/extrapolation geometries have separately regenerated SU2 8.0.1 data and full-field physical checks. The new fixed-architecture model has its own portable checkpoint and fitting/audit logs. Its finer CFD waveform/peak/drag errors are 7.2036%/2.2373%/2.2740%, with 15.3716% worst-case waveform error. Extrapolation waveform error is 34.2862%; this limitation is retained. Historical SU2 8.5.0 data, failed physical checks and model identities remain available as an instructional audit trail.

A separate unchanged historical checkpoint has 7.1499% aggregate waveform, 2.6343% peak and 2.0615% drag error against eight new physically checked finer CFD references. Its worst waveform error is 15.9701%. These numbers do not describe the new clean-label fit. The preserved original 8.5.0 test fields failed the added enthalpy allowance despite tiny residuals.
Detailed teaching guides explain geometry units, pressure normalization, data splitting, POD/scaler fitting, checkpoint inference, mesh sensitivity and claim boundaries. Two independent agent reviews addressed the NASA numerical evidence and the portable-model evaluation.

Scope and references

This is a reduced educational study, motivated by Zheng et al., Aerospace Science and Technology 178 (2026), 113218, DOI 10.1016/j.ast.2026.113218, and the distinct earlier article DOI 10.7638/kqdlxxb-2025.0081.

It does not reproduce the authors' full-aircraft CAD, 3,480-case dataset, neural weights, atmospheric propagation or ground-level PLdB. Those claims require additional author data and implementations. NASA CFD agreement does not establish experimental validity of a neural network trained on the two-parameter teaching body. The recovered finer-CFD neural arrays support numerical comparisons but do not restore missing original solver logs.

NASA sources and hashes are retained in cases/week16_lowboom/reference/source_manifest.json. See the Week 16 lecture, NASA reference guide, model validation guide and executable audit reports for definitions and limitations.

FlowMLLab v1.8.4 - Review follow-ups

FlowMLLab v1.8.4 - Review follow-ups

This maintenance release closes the small items left open by the independent
verification of v1.8.3. It changes no scientific result, no frozen evidence
archive and no historical prediction.

Week 15

• The six comparison figures carry one footer line again, and the bottom panel
  has its x/H axis label back (the v1.8.3 footer repair had painted over it).
  qa/repair_week15_figure_footers.py now starts from the v1.8.2 matplotlib
  renders and changes only the band below the bottom panel's tick labels.
• figure_provenance.json refreshed; the flagship notebook re-executed under a
  standard Jupyter kernel with the hash-verified Zenodo LR archive
  (notebook_execution.json updated); the 24-page lecture PDF rebuilt.
• results/week15_family_deeponet/README.md records that the seed-17 and
  seed-43 evidence was regenerated by the CI workflow on 19 September 2026, by
  how much it drifted (at most 1.4e-5 relative), and that the workflow no longer
  commits to main.

Documentation

• docs/GLOSSARY.md defines the acronyms and course terms; g005 is explained
  at its first use in the README; the Week 4.2 file numbering is noted.
• One labelling scheme for the Week 9 labs in the course map and lecture index,
  with a sentence explaining the lettered rows.
• Lecture section 35 names the repository files
  (results/week15_postaudit/notebook_execution.json, figure_provenance.json).
• START_HERE.md separates compute time from class time; the Week 12 fit time
  matches the week README. docs/RESULTS_GUIDE.md lists three Week 9 labs.
• PHYSICSNEMO_READINESS.md and PUBLISHING.md moved to docs/maintainers/.

Notebooks

• P6_FP_Cavity_Closure.ipynb stops with a clear message on a CPU-only machine
  at the CuPy check instead of failing ten cells later.

Verification

Fresh clone: pytest -q tests qa, flowmllab qa --root . and
python qa/validate_course_release.py pass; the Week 15 flagship notebook runs
with zero errors, 342 fields and six figure hashes verified; running every
notebook leaves the tracked tree unchanged. The wheel, source distribution,
lecture PDF and their SHA-256 list are attached to the GitHub release; the
Zenodo record for this version is created from .zenodo.json.

FlowMLLab v1.8.3 — verified release repair

FlowMLLab v1.8.3 — verified release repair

This release supersedes v1.8.2. It closes the remaining functional and release integrity defects found by a fresh-clone audit; it does not add new CFD or model training results.

Week 15

• Replaced the duplicated/clipped footer in all six comparison figures with one readable line and rebuilt the 24-page lecture PDF.
• Regenerated figure_provenance.json from the published PNG files.
• Re-executed the flagship notebook with the hash-verified Zenodo LR archive; all 15 code cells, 342 retained fields, and the six figure hashes pass.

Release integrity

• Synchronized pyproject.toml, the package version, CITATION.cff, README, results guide, release notes, and .zenodo.json at v1.8.3.
• Added ignore rules for documented notebook downloads and generated outputs.
• Evidence workflows now upload artifacts for review instead of committing regenerated binary evidence directly to main.
• The wheel and source distribution are built and checked, and this release carries both distributions, their hash list, and the Week 15 lecture PDF.

Verification

The tag was created only after GitHub Actions passed a fresh-clone real-kernel execution of the Week 15 flagship notebook, 242 tests (13 skipped), flowmllab qa, qa/validate_course_release.py, and a clean tracked checkout.

FlowMLLab v1.8.2 — repaired reproducibility release

This maintenance release supersedes v1.8.1, whose published tag contained five truncated notebook files. The scientific scope and frozen release data are unchanged; v1.8.2 repairs the affected teaching and reproducibility artifacts.

Repairs

• Restored intact, valid notebooks for Weeks 1, 3, 4, and both Week 15 activities.
• Restored the accompanying Week 15 lecture/evidence files required by those notebooks.
• Normalized the tracked Week 15 CSV to LF line endings so validation does not dirty a fresh checkout.

Verification

• 43 notebooks parsed as valid JSON.
• Fresh-clone test suite: 240 passed, 13 skipped (optional dependencies).
• flowmllab qa --root . and python qa/validate_course_release.py both passed.

The v1.8.0 evidence archives remain canonical on Zenodo: https://doi.org/10.5281/zenodo.22840293. This release corrects repository artifacts and does not claim new simulation, model-training, or experimental results.

FlowMLLab v1.8.1 — reproducibility and teaching corrections

This maintenance release implements the independent repository review of the
v1.8.0 Week 15 edition. It does not generate new CFD, retrain a model, alter a
frozen checkpoint, or replace a historical prediction.

Week 15 portability and scientific wording

• The flagship verification notebook now runs from a fresh clone plus the
  frozen Zenodo LR-sweep archive. Machine-local paths and an unnecessary
  PyTorch import were removed.
• The repository now links all four frozen Week 15 evidence archives and
  explains which packages have and have not received an independent audit.
• The three data-backed learning-rate and transfer figures used by the lecture
  are included with portable provenance paths.
• The lecture and README distinguish the historical, ordinary holdout and
  post-audit data splits; raw and filtered reverse-flow IoU; and the unresolved
  pressure nondimensionalization label.
• The SMART paper title, seed dependence at Re=25, the upper-edge learning-rate
  selection and the DoMINO velocity/IoU tradeoff are stated explicitly.
• The regenerated 24-page Week 15 PDF uses the corrected source and figures.

Course portability

• Week 7.4 downloads and hash-verifies its frozen release data when absent, so
  the Colab path no longer fails in its first data cell.
• Week 4 locates the repository independently of the notebook working
  directory, and the Week 7 LBM mathtext expression works with current
  Matplotlib.
• Student notebook runs write regenerated evidence to ignored student_runs/
  directories rather than overwriting retained release evidence.
• The release QA command prints a concise summary and writes its detailed JSON
  report under output/qa/.
• Version, notebook/PDF counts, page references, dependency descriptions and
  terminology were reconciled across the root documentation.

Verification

python -m pip install -e '.[test]'
python -m pytest -q tests qa
python -m flowmllab.cli smoke --root .
python qa/validate_course_release.py

The large v1.8.0 evidence archives remain canonical on
Zenodo; v1.8.1 changes the teaching
and verification layer around those frozen assets, not the assets themselves.

FlowMLLab v1.8.0 - Geometry-aware neural operators

FlowMLLab v1.8.0 - Geometry-aware neural operators

This release publishes the complete post-audit Week 15 geometry-generalization
edition. It builds on the public v1.7.0 wake-learning release and brings the
course to 43 notebooks and 26 lecture PDFs. Historical Week 15 runs remain
unchanged.

Week 15: geometry-aware neural operators

The new executed notebook and 24-page lecture compare ordinary DeepONet,
Geom-DeepONet, Geo-FNO, SMART, GeoTransolver and DoMINO, together with two
explicitly labelled company-inspired proxies. Six seven-row figures cover
Re=25, 50 and 100 at the physical 5:1 aspect ratio, with each prediction's own
streamlines and without the discarded orange threshold overlay.

The frozen 100/8/19 split is verified from geometry masks, g005 remains
quarantined, and no double-step motif enters training or validation. Because
the double-step family had already been inspected during the research program,
all results are described as retrospective rather than prospectively blind.
No new CFD was generated.

An equal 19,200-update ceiling corrects the earlier undertrained architecture
comparison. A non-double-step validation-only sweep selects peak learning rate
1e-3 for Geom, SMART and DoMINO. Historical Geom remains best in global
velocity error (9.86%); tuned Geom reaches 10.53%; and DoMINO has the largest
mean reverse-flow IoU in the tuned suite (0.534). The Re=25 low-step regime
remains the key limitation because its Reynolds-number/step-height combination
is absent from training. Negative ablations and failed vortex predictions are
retained.

Reproduction and evidence

python -m pip install -e '.[test,reconstruction]'
python -m pytest -q tests qa
flowmllab smoke --root .
flowmllab qa --root .
python qa/validate_course_release.py

The compact repository includes Week 15 summary tables, figure provenance and
executed teaching artifacts. Full checkpoints and saved prediction fields are
published as release/Zenodo evidence assets rather than committed to Git history.
Their names and SHA-256 digests are frozen in
release_assets_v1.8.0_sha256.txt.

The all-versions DOI is 10.5281/zenodo.22074169. The version-specific v1.8.0
DOI is 10.5281/zenodo.22840293.

FlowMLLab v1.6.1 - Teaching and reproducibility audit

FlowMLLab v1.6.1

This patch release promotes the repository-wide teaching and reproducibility audit completed after v1.6.0.

Highlights

• Improves definitions, notation, worked examples, and evidence interpretation across Weeks 1, 7.2, 9.3, 10.1, 11, 12, 14, and 15.
• Expands the Week 1 numerical-foundations lecture to a 27-page treatment of conservation laws, pressure elimination, streamfunction-vorticity coupling, finite differences, pressure recovery, stability, convergence, and validation.
• Rebuilds the Week 13 rectangular-cavity PINN notebook in teaching order with a real CPU training exercise and explicit separation of training loss, held-out residual, and CFD field error.
• Adds a staged Week 3 GPU cavity runner and repairs fresh-checkout paths and fallbacks in the beginner notebook sequence.
• Adds robust-scale Week 12 heat-flux figures while retaining the original full-range evidence.
• Removes private or upload-era context from public notebooks and documentation.
• Makes release QA index-driven, adds optional-dependency handling, and preserves failed scientific gates and limitations rather than relabelling them as successes.

Release verification

• 40 indexed notebooks and 24 indexed lecture PDFs.
• Package tests, scientific smoke tests, release QA, notebook HTML reproduction, and GitHub Pages deployment pass on the release candidate.
• The audited update contains 72 changed paths relative to v1.6.0-era main, including rebuilt PDFs and notebooks whose bytes were verified before merge.

The version-specific Zenodo DOI for v1.6.0 remains unchanged. The all-versions DOI is 10.5281/zenodo.22074169.

Week 7.4 diverse-wake dataset v1

Sixteen deterministic D2Q9-TRT trajectories for Week 7.4. The split and SHA-256 checksums are recorded in manifest.json.