DDSirota/reservoir-neural-bench

ReservoirNeuralBench — Norne dataset and OOD suite

Data companion of DDSirota/reservoir-neural-bench — a controlled benchmark of 20+ neural surrogates for 3-D reservoir simulation on the real Norne field geometry (46×112×22 corner-point grid, 44 431 active cells, OPM Flow ground truth).

No trained checkpoints are distributed here — train the released architectures yourself with the code in the GitHub repository (see Reproduce below).

Contents

path	size	what
norne_v1/	~4.1 GB	200 training/validation scenarios (5 warmup start states × 40 control schedules), 60 forecast steps × 30 days. Per scenario: states_active.npy (T, 44431, C) f16, controls (rates_actual.npz), metadata. Root: grid/static tensors, TPFA+NNC graphs, normalization stats (norne_norm_stats.json), audits.
norne_ood_v1/	~0.7 GB	30 held-out OOD control scenarios (well roles unchanged; only the rate schedule is OOD): far more skewed per-well rate allocation (interwell), field-rate + voidage-replacement intensity beyond training bounds (intensity), and a producer shut mid-run (toggle). Same grid/statics/normalization as norne_v1.

State-channel contract — 4 stored, 3 benchmarked. Each states_active.npy stores four dynamic fields — PRESSURE, SWAT, SGAS, RS (solution GOR). The published benchmark trains on the three-channel target PRESSURE / SWAT / SGAS (this is what the headline numbers and norne_norm_stats.json's meta.state_channels refer to); RS is shipped for extension work. See docs/NORNE_DATASET.md.

Download

pip install -U huggingface_hub
hf download DDSirota/reservoir-neural-bench --repo-type dataset --local-dir data/hf
# older CLI: huggingface-cli download DDSirota/reservoir-neural-bench --repo-type dataset --local-dir data/hf

Partial downloads work too, e.g. only the OOD suite:

hf download DDSirota/reservoir-neural-bench --repo-type dataset \
  --include "norne_ood_v1/*" --local-dir data/hf

Reproduce the headline numbers

No trained checkpoints are published — reproduce_headline.sh retrains the winner (well-gated multi-input TFNO, full-Nyquist, mixed target) for 3 seeds, then rolls it out. Needs a CUDA GPU; expect several GPU-hours total.

git clone https://github.com/DDSirota/reservoir-neural-bench
cd reservoir-neural-bench && pip install -r requirements.txt   # + neuraloperator, see requirements.txt
./reproduce_headline.sh data/hf

Expected (60-step autoregressive rollout, mean relL2 over PRESSURE/SWAT/SGAS at the final step): in-dist 0.0642 ± 0.0009, OOD 0.1071 ± 0.0049 over seeds 42/43/44; pressure RMSE 2.48 bar in-dist / 4.66 bar OOD. Small deviations from retraining are expected (GPU nondeterminism, library versions, hardware). Full tables, per-regime breakdown and the target-parameterization study: docs/RESULTS.md.

Provenance & license

Simulation outputs were generated with the open-source OPM Flow simulator on the Norne benchmark model distributed by the OPM project (open datasets, OPM/opm-data). OPM's opm-data notice applies the Open Database License (ODbL) v1.0 to decks/databases and the Database Contents License (DbCL) v1.0 to their contents unless stated otherwise. This published collection retains those upstream terms, notices, and attribution; no additional rights in the upstream Norne data are granted here. Terms for repository code are stated separately in the GitHub README. Full data attribution: DATA_LICENSE.md. Cite both this benchmark and the OPM/Norne data source.

Citation

@misc{reservoirneuralbench2026,
  author = {Sirota, Daniil},
  title  = {ReservoirNeuralBench: a controlled benchmark of neural surrogates for 3-D reservoir simulation},
  year   = {2026},
  url    = {https://github.com/DDSirota/reservoir-neural-bench}
}