Datasets:

DeCoDELab
/

GenISys

# Dataset guide

**220,999 design records · 189 public design parameters · three voyage ranges**

GenISys connects parametric bulkcarrier design with structural CAD, engineering
simulation, constraint evaluation, and generative models. This release contains
100,000 random designs, 117,999 repaired designs, and 3,000 SCDDPM-generated
attempts with required freight rate (RFR) guidance at **1,000, 5,000, and
8,000 nautical miles**.

The tables retain design inputs, engineering responses, constraint labels,
repair lineage, and evaluation failures. The random/repaired training release
and generated evaluation population are stored separately.

**Release status:** prepared and verified locally for manual upload to Hugging
Face. No new models, samples, physics evaluations, or CAD geometry were run to
assemble this release. The release license has not yet been specified.

## Populations

| Population | Designs | Completed evaluations | Verified feasible |
| --- | ---: | ---: | ---: |
| Random | 100,000 | 81,675 | 0 |
| Repaired | 117,999 | 88,764 | 16,247 |
| SCDDPM with RFR guidance | 3,000 | 2,998 | 1,793 |
| **Release total** | **220,999** | **173,437** | **18,040** |

![Population sizes and saved evaluation status](../figures/dataset_populations.png)

*Each bar uses all records in its population. The generated population comes
from a generator trained on feasible designs; these populations have different
construction and selection procedures.*

The unified random/repaired table contains 217,999 rows, including 47,560
failed evaluations. The generated population adds two failed evaluations.
A completed evaluation does not imply engineering feasibility. Verified
feasibility requires completion and a true `Global_V1p1_Feasible` label.

## SCDDPM designs with RFR guidance

These are the saved production attempts from the fixed-range RFR experiment
completed on Ada on 3 October 2026. Each range contains ten batches of 100
attempts from one saved feasible-only SCDDPM fit. All attempts are included;
there is no filtering for favorable performance or feasibility.

| Voyage range | Attempts | Completed | Feasible | Feasibility / all attempts | Mean feasible RFR (£/t) | Median feasible RFR (£/t) |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 1,000 nmi | 1,000 | 999 | 612 | 61.2% | 8.16 | 8.08 |
| 5,000 nmi | 1,000 | 1,000 | 593 | 59.3% | 18.81 | 18.56 |
| 8,000 nmi | 1,000 | 999 | 588 | 58.8% | 27.28 | 26.89 |

![Required freight rate distributions at the three voyage ranges](../figures/scddpm_rfr_distributions.png)

*Histograms show completed, feasible designs only. Dashed lines mark medians;
the annotations give feasibility against all 1,000 attempts per range. The
horizontal scales differ between panels. These are different mission lengths,
not matched-service comparisons. RFR uses the experiment's recorded economic
scenario in GBP per tonne.*

The saved run used 1,000 reverse-diffusion steps, feasibility guidance alpha
0.7, and RFR guidance alpha 0.1. Guidance began at timestep 799; voyage range
was fixed while the other independent controls were generated. The release
retains generator/critic checkpoint hashes and the original settings under
[`generated/scddpm_rfr_fixed_range/provenance/`](../generated/scddpm_rfr_fixed_range/provenance/).
These samples describe one saved generator fit, not independent retraining
replicates.

### Existing CAD examples from the paper

![Existing paper plate showing selected feasible generated bulkcarriers](../figures/paper_generated_design_examples.png)

*The paper's original CAD plate is reproduced unchanged. Its first row is an
RFR-guided SCDDPM validation design at approximately 1,981 nmi; the other two
are five-objective HCDDPM validation examples. These come from the paper's
separate variable-mission validation cohort and are illustrations of the design
representation. They are not the 1,000/5,000/8,000-nmi records released here.
The image uses independently fitted panel scales and display proxies; no new
geometry or engineering evaluation was generated for this README.*

## Files and directory structure

The release preserves the relative paths consumed by the code:

```text
Ship_Systems_Designs_v1p1/
  Dataset_100k/compiled/                         # Random population
  repair_experiments/Dataset_100k_Repaired/
    compiled/                                  # Repaired population
  Unified_Training_Dataset/                     # Shared training release
generated/scddpm_rfr_fixed_range/
  Generated_Designs.csv.gz                      # 3,000 joined sample records
  Range_Summary.csv                            # Saved-outcome summaries
  raw/range_{1000,5000,8000}/batch_00..09/       # Original batch files
  provenance/                                  # Settings, hashes, copy receipt
figures/                                       # README charts and paper plate
scripts/                                       # Repeatable packaging/figure script
file_manifest.json                             # Sizes and SHA-256 hashes
release_validation.json                        # Observed row/key/split counts
```

The base random/repaired release contains 44 copied data/support files.
The generated addition preserves 270 original batch/evaluation files, a joined
table, summaries, and provenance. Including the documentation figures, the
manifest covers approximately 2.59 GB. The unified table is a derived
representation of the random and repaired populations; its rows are not
additional designs.

The dataset card declares separate `unified`, `random`, `repaired`,
`split_assignments`, and `generated_scddpm_rfr` subsets using Hugging Face's
[manual configuration format](https://huggingface.co/docs/hub/datasets-manual-configuration).
The population subsets expose a `designs` split containing their complete population;
that display split does not replace the actual training/validation/test
assignments in `Shared_Split_Rows.csv.gz`. Other support tables remain
available as individual files. Hosted dataset-viewer rendering has not yet
been verified.

### Unified training release

| File | Role |
| --- | --- |
| `Training_Rows.csv.gz` | 217,999 rows and 424 columns: design inputs, identifiers, status, labels, and performance fields |
| `Shared_Split_Rows.csv.gz` | Frozen training/validation/test assignments |
| `Granular_20_Subsystem_Labels.csv.gz` | Labels used by the 20-subsystem model decomposition |
| `Granular_Subsystem_Labels.csv.gz` | Additional retained subsystem label representation |
| `Training_Manifest.json` | Population counts, feature lists, source paths, and training conventions |
| `Constraint_Subsystem_Mapping.csv` | Constraint-to-subsystem mapping |
| `Engineering_Constraint_Violation_Coverage.csv` | Recorded constraint violation coverage |
| `Constraints_Never_Violated.csv` | Constraints with no observed violations in this release |

Each design has 189 public parameters. The experiment schema identifies 133
independent learned controls and deterministically restores the remaining
coordinates. Parameter names, order, types, bounds, classes, and units are
specified in the source populations' `Parameter_Catalog.csv` tables and the
code repository's parameter schema.

### Source compiled tables

Each population includes design inputs, updated parameters, performance and
evaluation status, constraint catalogs, normalized constraint margins,
constraint bits, subsystem feasibility, and failure records. The repaired
population additionally includes `Repair_Lineage.csv.gz`.

Individual constraint margins are long-form tables. They are not one-row-per-
design training features. The random and repaired `ML_Design_Matrix.csv.gz`
files have different columns; use the unified release for shared training.

### Generated sample tables

`Generated_Designs.csv.gz` contains 3,000 rows and 400 columns. It joins the
original decoded 189-parameter vectors to their saved status and scalar
performance results using the original experiment metadata keys.

| Fields/files | Meaning |
| --- | --- |
| `Sample_ID` | Unique release key incorporating voyage range, batch, and sample position |
| `Range_nmi`, `Batch`, `Sample_In_Batch` | Original experiment identity |
| `Architecture`, `Study`, `Arm` | SCDDPM, fixed-range RFR study, RFR-guided arm |
| `Feasibility_Alpha`, `Performance_Alpha` | Recorded guidance settings |
| `Sampling_Seed` | Reconstructed from the archived configuration and runner; derivation is recorded in the release manifest |
| `Evaluator_Completed`, `Evaluation_Status` | Whether engineering evaluation completed |
| `Global_V1p1_Feasible`, `V1p1_Constraint_Fail_Count` | Saved feasibility and constraint outcomes |
| `Required_Freight_Rate_pounds_per_t` | Saved RFR in GBP per tonne |
| `raw/.../Generated_Controls_133.csv.gz` | Original model-generated independent controls |
| `raw/.../Submitted_Designs_189.csv.gz` | Original deterministic 133-to-189 decoded vectors |
| `raw/.../Postprocessing_Log.csv.gz` | Recorded deterministic decoding changes |
| `raw/.../Exact_Results.csv.gz` | Original status/performance table |
| `raw/.../exact_evaluation/compiled/` | Evaluator input vectors, constraints, performance, status, and failure evidence |

The decoded parameter vectors and the evaluator's recorded input vectors are
both retained under their original filenames. Use the compiled evaluator
vectors when a calculation requires the exact inputs after evaluator-side
canonicalization/cleanup. `Worker_Status == "completed"` means the worker
returned; it does not imply `Evaluator_Completed == True`.

## Identity, lineage, and splits

Use `(Population, Index)` as the unified design key. `Index` alone is not a
cross-population identifier. No duplicate population/index keys were found in
the unified table or shared split table.

Use `Sample_ID` for generated rows. Their original `Index`/`Generated_Index`
values repeat across batches and ranges. Do not join different generated
batches using those identifiers alone. Generated rows are an evaluation
population and are not appended to the frozen training table or split.

Preserve `Parent_Index`, `Structure_Index`, `Source_Structure_Index`, repair
lineage, and the released split assignments. Related source-parent designs
must not be independently reassigned across training and evaluation splits.

| Split | Rows |
| --- | ---: |
| Train | 152,347 |
| Validation | 32,637 |
| Test | 33,015 |

Status-aware eligibility and label availability are defined by the released
adapters and label tables. The random population has no globally feasible
designs; training a global feasibility classifier requires the appropriate
combined population and frozen split.

## Loading the data

Within the project's `GenISys_Env`:

```python
from pathlib import Path
import pandas as pd

release_root = Path("/path/to/downloaded/dataset")
unified = release_root / "Ship_Systems_Designs_v1p1/Unified_Training_Dataset"

rows = pd.read_csv(unified / "Training_Rows.csv.gz")
splits = pd.read_csv(unified / "Shared_Split_Rows.csv.gz")
labels = pd.read_csv(unified / "Granular_20_Subsystem_Labels.csv.gz")

generated = pd.read_csv(
    release_root / "generated/scddpm_rfr_fixed_range/Generated_Designs.csv.gz"
)
range_5000 = generated.loc[generated["Range_nmi"].eq(5000)]
feasible_5000 = range_5000.loc[
    range_5000["Evaluator_Completed"].eq(True)
    & range_5000["Global_V1p1_Feasible"].eq(True)
]
```

To use existing code defaults, place the downloaded
`Ship_Systems_Designs_v1p1/` directory under the GenISys code repository root.
That directory is excluded from Git tracking.

## Rebuilding this release and its figures

From an original GenISys workspace containing the saved experiment exports,
run with `GenISys_Env`:

```bash
python scripts/prepare_generated_samples_and_figures.py --source-root /path/to/GenISys
python scripts/validate_generated_samples.py
```

The script copies the approved production batches, verifies their hashes
against the original archive manifest, checks counts/RFR against the archived
arm summaries, builds the two charts, and copies the paper CAD plate unchanged.
It does not run a model, regenerate samples, rerun the evaluator, or build CAD.
PNG and SVG versions of the new charts are included.

The validation script checks the 270 raw-file hashes and compares every decoded
input and saved outcome in the joined table with its original batch record.

## Scope and limitations

These are synthetic designs evaluated by the GenISys early-design engineering
pipeline. Feasibility describes the implemented constraints and model scope;
it is not certification, class approval, or evidence that every regulatory
requirement is modeled. Numerical failures, method-domain limitations, missing
inputs, and physical constraint violations must retain their recorded status.

The dataset contains compiled tables and the selected generated batch files,
not the full per-design CAD/detail-output archive or trained model checkpoints.
Recreating CAD and evaluations requires
the corresponding GenISys code and environment. Source code and directories
marked `OLD` are not included in this dataset release.

## Provenance and reuse

Original files were copied without changing their contents. The original
random/repaired tables were compared to their sources using SHA-256. All 270
generated batch files also match the original archive's recorded hashes.
The joined generated table, summaries, and charts are derived documentation
artifacts with their own hashes; see `file_manifest.json` and the generated
population's `provenance/Release_Manifest.json`.

Population, identity, and split checks are recorded in `release_validation.json`.
Figure sources and chart counts are recorded in `figures/figure_manifest.json`.
Source manifests preserve their original recorded paths and dates as provenance.

Companion code: [GenISys on GitHub](https://github.com/noahbagz/GenISys).
Dataset: [GenISys on Hugging Face](https://huggingface.co/datasets/DeCoDELab/GenISys).
A release license and revision identifier have not yet been specified.