GeoProspect

A machine-learning pipeline that ranks undiscovered geothermal energy sites across the Western United States using real geophysical data.

---

The Question

Where in the Western US are the most promising undiscovered geothermal sites, and how confident are we?

The pipeline produces a calibrated probability score for every 0.1° grid cell (~11 km), SHAP explanations per site, and an interactive dashboard so a geologist can explore predictions and understand why a location scored high.

---

Project Structure

geoprospect/
├── app.py                         # Dash EDA dashboard (12-tab visual explorer)
├── requirements.txt
│
├── src/
│   ├── feature_engineering.py    # Builds feature matrix (0.1° grid × 11 features)
│   ├── train.py                  # XGBoost + spatial CV + SHAP + calibration + MLflow
│   ├── train_cnn.py              # 2D CNN on 5×5 geophysical patches + MLflow
│   └── download_known_sites.py   # One-time: downloads known geothermal site labels
│
├── data/
│   ├── heat/                     # SMU equilibrium HF + BHT CSVs
│   ├── gravity-geotiff/          # USGS Bouguer gravity XYZ
│   ├── Faults/                   # USGS Quaternary faults SHP + magnetic XYZ
│   ├── known_sites/              # NREL geothermal site labels (CSV + GeoJSON)
│   └── processed/
│       ├── feature_matrix.parquet    # Output of feature_engineering.py
│       ├── scored_grid.parquet       # Output of train.py (XGBoost scores + SHAP)
│       └── scored_grid_cnn.parquet   # Output of train_cnn.py (CNN scores)
│
├── models/
│   ├── xgb_model.pkl             # XGBoost model + calibrator
│   └── cnn_model.pt              # CNN weights + calibrator + normalization stats
│
└── mlruns/                       # MLflow experiment tracking (local)

---

Data Sources

All data is free and publicly available.

Dataset	Source	Records
Equilibrium Heat Flow	SMU via GDR/OpenEI	~8,400 filtered (15,655 raw)
BHT Heat Flow	SMU via GDR/OpenEI	~3,900 filtered (38,717 raw)
Bouguer Gravity	USGS mrdata	233,661 pts in WUS
Magnetic Anomaly	USGS mrdata	921,838 pts in WUS
Quaternary Faults	USGS Earthquake Hazards	109,670 segments (51,127 Holocene)
Known Geothermal Sites	NREL (operating + developing + low-temp)	1,553 WUS sites

Geographic scope: Western US — lon −125 to −102, lat 31 to 49.

---

Features

One row per 0.1° grid cell (~41,400 cells total). Aggregations use a 0.5° radius (~55 km).

Feature	Description	Source
hf_mean	Mean equilibrium heat flow within 0.5° (mW/m²)	SMU HF
hf_max	Max heat flow within 0.5°	SMU HF
hf_n	Number of HF measurements within 0.5°	SMU HF
bht_mean	Mean BHT-derived heat flow within 0.5°	SMU BHT
gravity_mean	Nearest Bouguer gravity value (mGal)	USGS Gravity
magnetics_mean	Nearest magnetic anomaly (nT)	USGS Magnetics
fault_length_total	Total fault length within 0.5° (km)	USGS Faults
fault_length_holocene	Holocene-equivalent fault length within 0.5° (km)	USGS Faults
dist_nearest_fault	Distance to nearest fault (km)	USGS Faults
dist_nearest_holocene	Distance to nearest Holocene-equivalent fault (km)	USGS Faults

Holocene definition: The USGS Quaternary fault database uses latest Quaternary (~last 15,000 yrs) and historic (~last 200 yrs) as the Holocene-equivalent age classes — not the string "Holocene". Feature engineering matches these explicitly.

Labeling: A cell is is_geothermal = 1 if a known site exists within 0.05° (~5.5 km). This tight radius (~985 positives, 2.4%) reflects the actual discovery problem — finding new sites — rather than regional classification.

dist_to_nearest_site is computed but never used as a model feature — only for filtering novel candidates post-inference.

---

Models

XGBoost

Gradient-boosted trees trained on the full 11-feature tabular matrix.

• Imbalance handling: scale_pos_weight = 41 (ratio of negatives to positives)
• Validation: 5-fold spatial GroupKFold with 3°×3° geographic blocks
• Calibration: Isotonic regression fit on out-of-fold predictions
• Explainability: SHAP values computed per cell via TreeExplainer
• Tracking: Logged to MLflow as xgb-spatial-cv

OOF performance: ROC-AUC = 0.753 | PR-AUC = 0.057 (random baseline = 0.024)

Top features by gain: hf_max (0.178) → dist_nearest_fault (0.155) → dist_nearest_holocene (0.099) → hf_mean (0.091)

CNN (2D Patch-Based)

A small 2D CNN that reads 5×5 spatial patches (~55 km wide) of 4 geophysical channels and predicts geothermal probability from local spatial context rather than point features.

Architecture: 2 conv layers (3×3) → BatchNorm → ReLU → Global Average Pooling → FC → sigmoid. ~5,361 parameters.

4 channels: hf_mean, gravity_mean, magnetics_mean, fault_length_total — channels with spatial structure. Distance/count features are excluded (already aggregated, no spatial pattern to convolve over).

Design choices:

• Reflection padding at map edges (avoids "ocean" zero-padding artifact)
• Per-channel z-score normalization (heat flow ~30–200, magnetics ~−1,600 to +2,700)
• Same 3°×3° spatial GroupKFold as XGBoost — fair comparison
• BCEWithLogitsLoss with pos_weight = 41 for imbalance
• Early stopping (patience = 8 epochs on val AP)
• Tracking: Logged to MLflow as cnn-spatial-cv

OOF performance: ROC-AUC = 0.761 | PR-AUC = 0.067 (random baseline = 0.024)

Ensemble

A cell-level mean of calibrated XGBoost and CNN scores — no separate training required. Computed at dashboard load time.

Top-1% prospect agreement:

• Both models flag: 395 cells (22.8%)
• XGBoost only: 1,338 cells
• CNN only: 820 cells

The low overlap reflects genuinely different inductive biases — XGBoost learns from point feature combinations while CNN learns from spatial neighbourhood patterns. The ensemble surfaces cells that both models agree on, yielding higher-confidence candidates.

Spatial Block Cross-Validation

Standard random k-fold leaks because adjacent cells share geology. The WUS is divided into 3°×3° geographic tiles; GroupKFold(n_splits=5) ensures no tile appears in both train and validation. All reported metrics are OOF (out-of-fold) — honest estimates of performance on unseen geographies.

---

Setup

python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -r requirements.txt

---

Running the Pipeline

Feature Engineering

python src/feature_engineering.py
# Runtime: ~3-5 min
# Output: data/processed/feature_matrix.parquet

Train XGBoost

python src/train.py
# Runtime: ~1-2 min
# Output: data/processed/scored_grid.parquet, models/xgb_model.pkl

Train CNN

python src/train_cnn.py
# Runtime: ~5-10 min (CPU) / ~1 min (GPU)
# Output: data/processed/scored_grid_cnn.parquet, models/cnn_model.pt

Dashboard

python app.py
# Open http://localhost:8050

MLflow Experiment Tracking

mlflow ui
# Open http://localhost:5000
# Experiment: geothermal-prospect-scoring
# Runs: xgb-spatial-cv, cnn-spatial-cv

---

Screenshots

Predictions	Faults

Heat Flow	Gravity

Feature Grid	Correlation

---

Dashboard (12 tabs)

Tab	What it shows
Heat Flow	SMU equilibrium HF measurements, quality filter, histogram
BHT	Bottom-hole temperature derived heat flow
Gravity	Bouguer gravity anomaly grid
Magnetics	Magnetic anomaly grid
Faults	Quaternary fault traces by age category (Holocene highlighted)
Known Sites	NREL operating, developing, low-temp sites
Overlay	All layers toggled on a single map
KDE Analysis	Geothermal sites vs random WUS background — separability check
Coverage	Data density per 0.2° cell (heat flow sparse, gravity dense)
Correlation	Feature correlation matrix + gravity vs heat flow scatter
Feature Grid	Any feature visualized as a map with geothermal site overlay
Predictions	XGBoost / CNN / Ensemble scores with threshold slider, novel-only filter

---

Output

scored_grid.parquet (XGBoost):

Column	Description
score	Calibrated geothermal probability
score_raw	Uncalibrated XGBoost output
is_novel	1 if > 50 km from any known site
shap_*	SHAP attribution per feature

scored_grid_cnn.parquet (CNN):

Column	Description
cnn_score	Calibrated CNN probability
cnn_score_raw	Uncalibrated sigmoid output
is_novel	1 if > 50 km from any known site

Novel candidates = cells where is_novel == 1 AND is_geothermal == 0, ranked by score. These are the actionable predictions for follow-up exploration.

---

Geological Interpretation

Signal	Geothermal meaning
High hf_max	Direct evidence of elevated crustal heat flux
Low dist_nearest_fault	Proximity to fracture network → fluid pathway
Low dist_nearest_holocene	Active fault → permeable, likely fluid-bearing zone
High hf_mean	Sustained regional heat anomaly
Negative gravity_mean	Low-density, thermally expanded crust (Basin & Range)

The Basin & Range province (Nevada, western Utah, Arizona) dominates high-scoring regions because it combines all five signals.

---

Known Limitations

• Heat flow has 28% cell-level missingness (sparse measurement coverage). Imputed with column median before training.
• Gravity and magnetics features are nearest-neighbour lookups (dense grids), not kriged surfaces.
• The "negative" class contains unmapped geothermal sites — PU (Positive-Unlabeled) learning would be more rigorous than scale_pos_weight.
• CNN uses only 4 channels; dist_nearest_fault and count features are excluded because they don't have spatial raster structure to convolve over.
• Fold 2 CNN PR-AUC (0.049) is notably weaker than other folds — that geographic block has fewer positives (155) and likely less signal overlap with training regions.