SiliciclasticReservoirs

1,000,000 synthetic 3D siliciclastic-reservoir geology cubes generated from rule-based sedimentological simulations (turbidite lobes + 6 fluvial-channel architectures + delta-fan distributary). Cubes are voxelized at (64, 64, 32) cells. Each sample carries facies, porosity, permeability, and a structured set of geological conditioning parameters.

Designed to train conditional generative models (flow matching, diffusion, etc.) of subsurface geology under interpretable physical conditioning.

Quick stats

• 1,000,000 samples across 8 reservoir architectures
• Cube shape: (64, 64, 32) voxels — (x, y, z), z is depth (0 = base)
• 4 voxel arrays per sample: binary facies, 6-class facies, porosity, permeability
• Compact slim parquet (9–12 cols) and full reproducibility parquet (25–75 cols)
• Captions for text-conditioning experiments
• Fully reproducible — every sample regenerable from (seed, params)

File structure

SiliciclasticReservoirs/
├── README.md                   ← this file
├── DATASHEET.md                ← Datasheet for Datasets (Gebru et al. 2018)
├── splits/                     ← deterministic 90/5/5 split, stratified by layer_type
│   ├── train.parquet           ← 900,000 rows
│   ├── validation.parquet      ←  50,000 rows
│   └── test.parquet            ←  50,000 rows
├── lobe/                       ← 200,000 samples
│   ├── shard_0000/
│   │   ├── facies.npy          (n, 64, 64, 32) int8 — binary 0/1
│   │   ├── facies_alluvsim.npy (n, 64, 64, 32) int8 — 6-class -1..4
│   │   ├── poro.npy            (n, 64, 64, 32) float16 — porosity in [0, 0.5]
│   │   ├── perm.npy            (n, 64, 64, 32) float16 — permeability in mD [0, 60000]
│   │   ├── params_slim.parquet ← training conditioning (use this)
│   │   └── params.parquet      ← full physics for reproducibility
│   ├── shard_0001/
│   └── … 256 shards total
├── channel_pv_shoestring/      ← 100,000 samples, 256 shards
├── channel_cb_labyrinth/       ← 100,000 samples, 256 shards
├── channel_cb_jigsaw/          ← 150,000 samples, 256 shards
├── channel_sh_distal/          ← 100,000 samples, 256 shards
├── channel_sh_proximal/        ← 100,000 samples, 256 shards
├── channel_meander_oxbow/      ← 100,000 samples, 256 shards
└── delta/                      ← 150,000 samples, 256 shards

256 shards per preset, each ~380 MB – 1.2 GB depending on per-shard sample count.

Layer types

These follow Pyrcz & Deutsch's standard fluvial-architecture taxonomy ("User Guide to the Alluvsim Program", 2004).

Voxel arrays per sample

All four arrays are row-aligned within a shard (sample i of facies corresponds to row i of the parquets and to sample i of all other arrays).

Slim parquet (`params_slim.parquet`) — the training conditioning

This is what you train on as conditioning. It carries the geological parameters that vary across samples.

Universal columns (all 8 layer types)

Family-specific columns (NULL outside family)

Total: 9 cols for lobe rows, 11 for channel rows, 12 for delta rows. Rows from different families share the universal columns and have NULL in non-applicable family columns.

Full parquet (`params.parquet`) — for reproducibility

25 columns for lobe rows, 73 for channel rows, 75 for delta rows. Includes the per-realization seeds, the per-event statistical knob widths (stdev*), the engine-internal event-budget knobs (ntime*), and the FACIES_PROPS extras. Don't train on this — most columns are constant or redundant. Useful for:

• Re-generating a specific sample bit-for-bit (ChannelLayer(...).create_geology(**row.to_dict()) with the saved seed)
• Diagnostic analysis of the engine's behavior

Splits

splits/{train,validation,test}.parquet define a deterministic 90 / 5 / 5 partition stratified by layer_type (each split has all 8 architectures in proportion). Generated with seed 42.

Each row is (layer_type, shard_dir, sample_idx) pointing into one of the data shards.

Cell-physical-size note

Cube shape is identical (64×64×32) but the physical cell size differs by family:

For training in cell-units, this difference is hidden inside width_cells / depth_cells. The model can ignore physical scale and just train on the cube as-is.

Loading example (PyTorch)

import numpy as np
import pyarrow.parquet as pq
from pathlib import Path
import torch
from torch.utils.data import IterableDataset

class ReservoirShardDataset(IterableDataset):
    """One worker walks an assigned subset of shards, yields (cube, params)."""
    def __init__(self, root: Path, split: str = "train"):
        self.root = Path(root)
        # Read split index
        self.index = pq.read_table(self.root / f"splits/{split}.parquet").to_pylist()

    def __iter__(self):
        worker = torch.utils.data.get_worker_info()
        my_rows = (self.index if worker is None
                   else self.index[worker.id::worker.num_workers])
        cur_shard = None; cur_data = None
        for row in my_rows:
            shard_dir = self.root / row["shard_dir"]
            if shard_dir != cur_shard:
                cur_shard = shard_dir
                cur_data = {
                    "facies": np.load(shard_dir / "facies.npy", mmap_mode="r"),
                    "poro":   np.load(shard_dir / "poro.npy",   mmap_mode="r"),
                    "perm":   np.load(shard_dir / "perm.npy",   mmap_mode="r"),
                    "params": pq.read_table(
                        shard_dir / "params_slim.parquet").to_pylist(),
                }
            i = row["sample_idx"]
            yield {
                "facies": np.asarray(cur_data["facies"][i], dtype=np.int8),
                "poro":   np.asarray(cur_data["poro"][i],   dtype=np.float32),
                "perm":   np.asarray(cur_data["perm"][i],   dtype=np.float32),
                "params": cur_data["params"][i],
            }

ds = ReservoirShardDataset("/path/to/SiliciclasticReservoirs", split="train")
loader = torch.utils.data.DataLoader(ds, batch_size=8, num_workers=4)

Recommended preprocessing for flow-matching training

• facies → [0, 1] float: just cube.astype(np.float32)
• poro → standardized: (poro - 0.15) / 0.10 keeps it roughly N(0, 1)-ish for sand cells
• perm → log-perm: np.log10(np.maximum(perm, 1e-3)), then maybe standardize
• layer_type → categorical embedding (8 classes). Use layer_type.split(":")[0] for family-level (3 classes) or full string for fine-grained (8 classes).
• Other slim columns → continuous embeddings, with null → learned sentinel or per-family mask.

Geological caveats

• Channel topology is tree-like, not anastomosing. Avulsions split channels but never re-merge. The braided / jigsaw look in CB_LABYRINTH / CB_JIGSAW comes from heavily overlapping stamps in the cube, not from explicit topological merging. For facies-based generative training this doesn't matter; for downstream flow simulation, be aware.
• Per-cell poro/perm follow a Walker-1992 upward-fining ramp (highest at thalweg base, decaying upward and laterally toward banks) plus a per-event Kozeny-Carman-coupled draw plus a per-realization "regional rock quality" multiplier. Channels in the same cube vary in quality (different events had different sediment supply); different cubes span the full quality range.
• Channels have FACIES_PROPS-driven base poro/perm values that are conservative midpoints from fluvial-reservoir literature (Pyrcz & Deutsch 2002, Allen 1965, Walker 1992); they're not from a specific reservoir analog.
• Delta has no mouth bars painted (paint_mouth_bars=False). The delta is a distributary network, not a clinoform with bar deposits.
• Cubes are post-cropped (8 cells trimmed from each lateral edge, 9 from top + 9 from bottom of z) to remove engine boundary artifacts. The internal engine ran on a slightly larger grid before cropping.

See DATASHEET.md for full documentation.

Citation

@misc{SiliciclasticReservoirs_2026,
  author       = {Anonymous},
  title        = {{SiliciclasticReservoirs}: 1M Synthetic 3D Reservoir Geology Cubes for Conditional Generative Modeling},
  year         = {2026},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/AnonymouScientist/SiliciclasticReservoirs}}
}

License

Creative Commons Attribution 4.0 International (CC-BY-4.0). Use, share, modify, redistribute — just attribute the source.

Contact / Issues

File issues at the dataset repository on HuggingFace.

📊 Structured Schema (Zero-Fabrication)

Estimated Rows: 1,000,000