AI-Driven Reservoir Performance Proxy

Physics-Aware Deep Learning | ML2 Final Project — University of Chicago

Replace months of reservoir simulation with millisecond ML inference. Given 6 geological and operational parameters, our models predict 22-year oil production, cumulative output, and reservoir pressure with R² > 0.94 across all targets.

The Problem

Reservoir simulators (like OPM Flow) solve complex fluid-flow equations over millions of grid cells. A single 22-year scenario takes minutes-to-hours to run — making real-time decision-making, history matching, and economic optimization impractical.

Our solution: train surrogate models on 200 high-fidelity OPM Flow simulations so that the same prediction takes under 1 second.

Models

Model	Type	Predicts	Best R²
Encoder-Decoder LSTM	Deep learning	Full 22-year time-series (8 variables)	0.991 †
MLP	Deep learning	4 final/peak scalar values	0.996
PINN	Physics-informed NN	4 scalars with Darcy's Law enforced	0.996
Random Forest	Classical ML	4 scalars (most interpretable)	0.988

† Best single-variable R² (FGIT — cumulative gas injection); average R² across all 8 output variables is 0.94.

Input Features (6 parameters)

Parameter	Description	Example range
`producer_bhp_psi`	Producer bottom-hole pressure	1000–5000 psi
`gas_inj_rate_mscf_d`	Gas injection rate	5–100 Mscf/d
`inj_bhp_limit_psi`	Injector BHP ceiling	5000–15000 psi
`init_prod_period_days`	Initial production period	100–5000 days
`perm_multiplier`	Permeability multiplier	0.1–10.0
`poro_multiplier`	Porosity multiplier	0.5–2.0

Output Targets

Variable	Description	Unit
`fopr` / `fopt`	Oil production rate / Cumulative oil	STB/d, STB
`fpr`	Field reservoir pressure	PSIA
`fgpr` / `fgpt`	Gas production rate / Cumulative gas	MSCF/D, MSCF
`fgir` / `fgit`	Gas injection rate / Cumulative injection	MSCF/D, MSCF
`wbhp_inj`	Injector bottom-hole pressure	PSIA

Quickstart

1. Install dependencies

pip install -r requirements.txt

2a. Try the live demo (no installation needed)

Run predictions directly in your browser — no code required: 👉 huggingface.co/spaces/mushahid-raza/reservoir-proxy-demo

2b. Download pre-trained weights (recommended)

Skip retraining entirely — download all saved models with one command:

pip install huggingface_hub
python download_models.py

This fetches all .keras, .pkl, .npy, and .json files from HuggingFace into saved_models/.

2c. Or retrain from scratch

Run each notebook in order. Each saves trained models to saved_models/:

04_random_forest_baseline.ipynb   → saved_models/rf_*.pkl
02_mlp_proxy.ipynb                → saved_models/mlp_*.keras / .pkl
03_pinn_proxy.ipynb               → saved_models/pinn_*.keras / .pkl
01_lstm_encoder_decoder.ipynb     → saved_models/enc_dec_lstm_*.keras / .pkl / .npy

3. Run inference — command line

# Full 22-year time-series from LSTM
python predict.py --model lstm \
  --perm 1.5 --poro 1.2 --bhp 2000 \
  --inj_rate 50 --inj_bhp 8000 --init_period 365 --plot

# Compare all 4 models on scalar targets
python predict.py --model all \
  --perm 1.0 --poro 1.0 --bhp 3000 \
  --inj_rate 35 --inj_bhp 10000 --init_period 3230

4. Run inference — interactive notebook

Open demo.ipynb, edit the 6 parameters in Cell 1, and run all cells.

Repository Structure

├── 01_lstm_encoder_decoder.ipynb       # Encoder-Decoder LSTM (full time-series)
├── 02_mlp_proxy.ipynb                  # MLP scalar proxy
├── 03_pinn_proxy.ipynb                 # Physics-Informed NN
├── 04_random_forest_baseline.ipynb     # Random Forest baseline
├── demo.ipynb                          # Interactive inference demo
├── predict.py                          # CLI inference script
├── download_models.py                  # Fetch weights from HuggingFace
├── requirements.txt                    # Python dependencies
├── dataset_scalar.csv                  # Dataset for MLP / PINN / RF
├── dataset_timeseries_lstm.csv         # Dataset for LSTM
├── figures/                            # Training and evaluation plots
│   ├── lstm_pred_vs_actual.png
│   ├── rf_feature_importance.png
│   ├── rf_pred_vs_actual.png
│   └── rf_residuals_by_group.png
└── saved_models/                       # Created after running notebooks
    ├── enc_dec_lstm_reservoir_proxy.keras
    ├── scaler_static.pkl / scaler_time.pkl / scaler_y.pkl
    ├── lstm_avg_time_grid.npy
    ├── mlp_reservoir_proxy.keras
    ├── mlp_scaler_X.pkl / mlp_scaler_y.pkl
    ├── pinn_base_model.keras
    ├── pinn_scaler_X.pkl / pinn_scaler_y.pkl
    └── rf_final_fopt.pkl / rf_final_fpr.pkl / rf_final_fopr.pkl / rf_peak_fopr.pkl

How the Dataset Was Generated

We used the SPE10 benchmark reservoir model as our base simulation. Across 6 key parameters, we applied Latin Hypercube Sampling to generate 200 diverse scenarios:

Runs 1–100: Operational sensitivity (isolating individual variable effects)
Runs 101–150: Chaotic group (simultaneous variation for interaction capture)
Runs 151–200: Geological group (permeability/porosity uncertainty)

Each scenario was simulated in OPM Flow producing 22 years of production data (~800 adaptive time steps per run → 158,742 total rows).

Key Technical Contributions

1. Encoder-Decoder LSTM with sentinel padding Static reservoir parameters are compressed into a latent "scenario fingerprint" by the Encoder; the Decoder LSTM unrolls it over adaptive time steps. Padded sequences use a −1.0 sentinel with a Keras Masking layer to prevent the LSTM from learning on padded steps.

2. Weighted Huber loss for integral drift Standard MSE causes cumulative oil (FOPT) to drift because it doesn't enforce that cumulative = ∫ rate. We apply a 1.5× loss weight to FOPT/FGPT to correct this — R²(FOPT) improved from 0.57 to 0.965.

3. Physics-Informed Neural Network Two reservoir engineering constraints are embedded directly in the training loop via a custom GradientTape:

Darcy's Law: FPR > Producer BHP (flow cannot reverse)
Injection ceiling: FPR < Injector BHP limit

Violations are penalized with Huber loss (λ = 0.01), achieving 93% physical constraint compliance (56/60 checks passed) on the test set and R²(FOPT) = 0.996.

Results

Target	LSTM R²	MLP R²	PINN R²	RF R²
Oil Rate (FOPR)	0.984	0.984	0.942	0.972
Cumul. Oil (FOPT)	0.965	0.996	0.996	0.988
Reservoir Pressure (FPR)	0.842	0.972	0.960	0.934
Gas Injection Rate (FGIR)	0.947	—	—	—
Cumul. Gas Injection (FGIT)	0.991	—	—	—

Note on LSTM FPR (R² = 0.842): Reservoir pressure exhibits slower convergence in recurrent models due to pressure equilibration dynamics spanning the full 22-year horizon. For pressure-critical predictions, use the MLP (R² = 0.972) or PINN (R² = 0.960) scalar models instead.

LSTM: Predicted vs Actual (3 held-out test scenarios)

Authors

Sabayna Ali · Gabe Horas · Morgan Klutzke · Mushahid Raza
Machine Learning 2 — University of Chicago, Spring 2026

Citation

If you use this work, please cite:

BibTeX:

@misc{ali2026reservoir,
  title     = {AI-Driven Reservoir Performance Proxy: Physics-Aware Deep Learning for Surrogate Reservoir Modeling},
  author    = {Ali, Sabayna and Horas, Gabe and Klutzke, Morgan and Raza, Mushahid},
  year      = {2026},
  publisher = {University of Chicago},
  note      = {ML2 Final Project. Models: \url{https://huggingface.co/mushahid-raza/reservoir-proxy-models}},
}

Plain text:

Ali, S., Horas, G., Klutzke, M., & Raza, M. (2026).
AI-Driven Reservoir Performance Proxy: Physics-Aware Deep Learning
for Surrogate Reservoir Modeling.
University of Chicago.

Downloads last month: 144

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

mushahid-raza
/

reservoir-proxy-models