ddsirota/reservoir-neural-bench

How ddsirota/reservoir-neural-bench works

ReservoirNeuralBench benchmarks neural surrogates for one-step, three-phase Norne reservoir forecasts and evaluates them through 60-step autoregressive rollouts, including out-of-distribution well controls. The product workflow runs from dataset preparation and actual-rate inputs through model training to rollout validation and reporting. Its model portfolio spans spectral, convolutional, operator-learning, and graph architectures. The README documents the benchmark stages and capabilities; source excerpts and the SOURCE INDEX support selected implementation links, so unsampled subsystem connections are not assumed.

Main entry points: train_norne.py Model training · norne_generate_dataset.py Dataset generation · validate_rollout_norne.py Rollout validation · make_report_norne.py Benchmark report

Components 28

Norne data

Dataset generation tools/norne_generate_dataset.py
Realized rates tools/norne_extract_actual_rates.py
Well features tools/norne_extract_well_index.py
Graph construction tools/norne_build_graph.py
Dataset export tools/norne_export_dataset.py
Norne loader data/norne_loader.py

Surrogate models

TFNO models/tfno.py
Multi-input Fourier models/fourier_mionet.py
U-FNO models/u_fno.py
DeepONet models/deeponet.py
MIONet variants models/mionet_deeponet.py
DeepXDE branches models/deepxde_branch_deeponet.py
Message passing GNN models/gnn.py
MeshGraphNet models/meshgraphnet.py
Multiscale graph net models/mgn.py
Flux GNN models/flux_gnn.py
TPFA GNN models/tpfa_gnn.py
3D U-Net models/unet3d.py
Residual U-Net models/resunet3d_hdc.py
Conv3D residual models/residual.py

Training and evaluation

Model training train_norne.py
Rollout validation validate_rollout_norne.py
Mass balance tools/mass_balance.py
Fluid properties tools/norne_fluid_props.py

Benchmark analysis

Benchmark report make_report_norne.py
Figures tools/make_figures.py
Rollout error curves tools/make_rollout_error_curves.py

Other components

Benchmark user

Connections 29

Benchmark user → Dataset generation: generate data README.md
Dataset generation → Realized rates: extract rates
Dataset generation → Well features: derive well features
Dataset generation → Graph construction: build graph
Dataset generation → Dataset export: export dataset
Benchmark user → Model training: train models README.md
Model training → Norne loader: loads batches train_norne.py
Model training → TFNO: uses model registry train_norne.py
Model training → Multi-input Fourier: selects model
Multi-input Fourier → TFNO: uses backbone models/fourier_mionet.py
Benchmark user → Rollout validation: validate forecasts README.md
Rollout validation → Norne loader: loads scenarios validate_rollout_norne.py
Rollout validation → Mass balance: checks balance validate_rollout_norne.py
Rollout validation → Fluid properties: uses fluid properties validate_rollout_norne.py
Benchmark user → Benchmark report: review results README.md
Benchmark report → Figures: summarizes benchmark
Benchmark report → Rollout error curves: presents rollout errors
Model training → Message passing GNN: selects model train_norne.py
Model training → MeshGraphNet: selects model train_norne.py
Model training → Multiscale graph net: selects model train_norne.py
Model training → Flux GNN: selects model train_norne.py
Model training → TPFA GNN: selects model train_norne.py
Model training → DeepONet: selects model
Model training → MIONet variants: selects model
Model training → DeepXDE branches: selects model
Model training → 3D U-Net: selects model
Model training → Residual U-Net: selects model
Model training → Conv3D residual: uses component train_norne.py
Model training → U-FNO: selects model

Drawn by GitDiagram on November 13, 2026 from the repository's file tree, README and source files. Read as Markdown ·