GeoForce — CNN Geothermal Reservoir Surrogate¶
Status: Deployed v1.1 Platform: platform.forcex-ai.com Model: Published on HuggingFace (GeoForce-CNN-v1.1) Dataset: Published on HuggingFace (geothermal-reservoir-dataset)
Executive Summary¶
A CNN-based surrogate model that replaces hours of TOUGH2 geothermal reservoir simulation with sub-second inference — achieving R²=0.997 accuracy with only 57K parameters. GeoForce is the flagship deployed product of the ForceX AI platform, bridging the gap between physics simulation fidelity and real-time operational decision-making.
Key Results¶
| Metric | Value |
|---|---|
| Model Accuracy | R² = 0.997 |
| Parameters | 57K (lightweight) |
| Inference Time | Sub-second (vs hours for TOUGH2) |
| Training Data | Synthetic TOUGH2 simulation outputs |
| Version | v1.1 (improved from initial ForceX prototype) |
Technical Architecture¶
Model Design¶
- Architecture: Convolutional Neural Network optimized for spatial reservoir data
- Input: Reservoir parameters (permeability, porosity, injection rates, boundary conditions)
- Output: Temperature and pressure field predictions across the reservoir grid
- Training: Supervised learning on TOUGH2 simulation outputs
- Validation: Cross-validated against held-out simulation scenarios
Physics-Informed Approach¶
GeoForce maintains physical consistency by:
- Training on physics-simulator output (TOUGH2) ensuring conservation laws are respected
- Encoding spatial relationships through convolutional layers
- Validating predictions against known geothermal field behavior
- Monitoring physical plausibility metrics alongside ML accuracy
Technology Stack¶
| Component | Technology |
|---|---|
| ML Framework | PyTorch |
| Simulation Engine | TOUGH2 (data generation) |
| API | FastAPI |
| Frontend | React + Vite + TypeScript (dashboard) |
| Deployment | Docker, VPS |
| Model Registry | HuggingFace Hub |
| Dataset Hosting | HuggingFace Datasets |
| Testing | pytest |
Repository Structure¶
GeoForce/
├── surrogate/ # CNN model architecture and training
├── simulation/ # TOUGH2 simulation interface
├── solver/ # Physics solver utilities
├── agent/ # LangGraph agent for interactive queries
├── app/ # FastAPI backend
├── dashboard/ # React visualization dashboard
├── data/ # Training and validation datasets
├── notebooks/ # Exploration and analysis notebooks
├── demo/ # Demo scripts
├── tests/ # Test suite
├── docs/ # Documentation
└── tools/ # Utility scripts
Evolution from ForceX AI¶
GeoForce started as one of 12 products inside the ForceX AI platform. The standalone repository represents the production-hardened version with:
- Improved CNN architecture (v1.0 → v1.1)
- Published model weights on HuggingFace for reproducibility
- Published training dataset for community use
- Dedicated LangGraph agent for interactive reservoir analysis
- Production API with monitoring and dashboards
Use Cases¶
- Geothermal Field Development: Rapid scenario analysis for well placement
- Operational Optimization: Real-time reservoir monitoring and prediction
- Exploration: Quick screening of potential geothermal sites
- Education: Teaching reservoir physics with instant feedback
Skills Demonstrated¶
- CNN Surrogate Modeling
- Physics-Informed Machine Learning
- Geothermal Reservoir Simulation (TOUGH2)
- Model Publishing (HuggingFace Hub)
- Dataset Curation and Publishing
- Production API Deployment