GeoForce-CNN-v1.1

GeoForce-CNN v1.1: Physics-Informed CNN Surrogate for Geothermal Reservoir Prediction

GeoForce is a physics-informed convolutional neural network (CNN) that predicts temperature and pressure field evolution in geothermal reservoirs. It replaces computationally expensive numerical simulators with a compact 57,802-parameter model that runs in 3.2ms on CPU — a 282,350x speedup over finite-difference simulation.

<div align="center"> </div>

Model Description

GeoForce maps static reservoir properties (temperature, permeability, porosity, depth, well locations, pressure) directly to spatiotemporal temperature and pressure fields over a 20-year production horizon. It was designed for rapid uncertainty quantification and well placement optimization in Indonesian geothermal systems.

Architecture

Encoder-Residual-Decoder CNN:

Input (batch, 6, 32, 32)
    ↓
[Encoder]
    ConvBlock: Conv2d(6→32, 3×3) + BatchNorm + ReLU
    ConvBlock: Conv2d(32→32, 3×3) + BatchNorm + ReLU
    ↓
[Middle]
    ResidualBlock: Conv→BN→ReLU→Conv→BN + skip → ReLU
    ResidualBlock: Conv→BN→ReLU→Conv→BN + skip → ReLU
    ↓
[Decoder]
    ConvBlock: Conv2d(32→32, 3×3) + BatchNorm + ReLU
    Conv2d(32→10, 1×1) + Sigmoid
    ↓
Output (batch, 10, 32, 32)
  → Channels 0-4: Temperature at years 4, 8, 12, 16, 20
  → Channels 5-9: Pressure at years 4, 8, 12, 16, 20

Input Channels (6)

Physics-Informed Loss

The training loss combines data-driven MSE with three physics constraints:

L_total = L_data + 0.1 × L_physics
L_physics = 0.1 × L_temporal + 0.01 × L_spatial + 0.001 × L_darcy

• Temporal smoothness: Penalizes non-physical jumps between timesteps (energy conservation proxy)
• Spatial smoothness: Discrete Laplacian penalty enforcing diffusion behavior
• Darcy coupling: Constrains pressure gradients in high-permeability zones

Training

• Data: 1,000 synthetic reservoir simulations via implicit finite-difference solver (coupled heat + Darcy flow)
• Parameter ranges: Calibrated to Indonesian geothermal fields (Kamojang, Wayang Windu, Darajat)
• Split: 800 train / 100 val / 100 test (seed=42)
• Optimizer: Adam (lr=1e-3, ReduceLROnPlateau)
• Training time: 44.9 minutes on CPU (no GPU required)
• Best epoch: 352 / 402 (early stopped, patience=50)
• Best validation loss: 0.000382

Quick Start

import torch
import numpy as np
from reservoir_cnn import ReservoirCNN

# Load model
model = ReservoirCNN(in_channels=6, n_time_steps=5, base_filters=32)
checkpoint = torch.load("geoforce_cnn_v1.1.pt", map_location="cpu", weights_only=False)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Prepare input: 6-channel tensor on 32×32 grid
# Example: Kamojang-like reservoir
base_temp = 245.0     # °C
log_perm = -14.0      # log10(m²)
base_pressure = 15e6  # Pa (15 MPa)
porosity = 0.08
depth = 1500.0        # m
n_wells = 3

# Normalize inputs
inp = torch.zeros(1, 6, 32, 32)
inp[0, 0] = (base_temp - 25) / 325          # Temperature
inp[0, 1] = (log_perm + 16) / 4             # Permeability
# Channel 2: well mask (place wells with Gaussian decay)
for _ in range(n_wells):
    wx, wy = np.random.randint(4, 28, size=2)
    for i in range(32):
        for j in range(32):
            d = ((i - wy)**2 + (j - wx)**2) / 50.0
            inp[0, 2, i, j] = max(inp[0, 2, i, j].item(), np.exp(-d))
inp[0, 3] = (base_pressure - 5e6) / 20e6    # Pressure
inp[0, 4] = (porosity - 0.01) / 0.14        # Porosity
inp[0, 5] = (depth - 800) / 1700            # Depth

# Run inference
with torch.no_grad():
    output = model(inp)  # (1, 10, 32, 32) normalized [0,1]

# Denormalize to physical units
temperature = output[0, :5] * 350.0          # °C, shape (5, 32, 32)
pressure = output[0, 5:] * 50e6              # Pa, shape (5, 32, 32)

print(f"Year 20 avg temperature: {temperature[4].mean():.1f}°C")
print(f"Year 20 avg pressure: {pressure[4].mean()/1e6:.2f} MPa")

Validation Results

Per-Timestep Accuracy

Benchmark Comparison

Version History

The v1.0 failure was caused by compressing three physical parameters (pressure, porosity, depth) into a single input channel. Each parameter now has its own dedicated channel in v1.1. Full failure analysis is documented in the technical report.

Limitations

1. 2D grid only (32×32 horizontal slice) — no vertical flow or gravity convection
2. Single-phase liquid water — no steam-water phase transitions
3. Homogeneous permeability per scenario — no fracture networks
4. Synthetic training data — not yet validated against real field measurements
5. Max error of 47.3°C in extreme edge cases near training distribution boundaries

Use Cases

• Uncertainty quantification: Run 10,000+ Monte Carlo samples in < 1 minute
• Well placement optimization: Evaluate hundreds of configurations in seconds
• Screening studies: Rapidly assess reservoir viability before committing to full simulation
• Education: Demonstrate reservoir physics concepts with instant feedback

Files

Citation

@software{geoforce2026,
  title={GeoForce: Physics-Informed CNN Surrogate for Geothermal Reservoir Prediction},
  author={Riupassa, Robi Dany},
  year={2026},
  publisher={ForceX AI},
  url={https://huggingface.co/ForceX-AI/GeoForce-CNN-v1.1},
  version={1.1.0}
}

About ForceX AI

ForceX AI builds AI-powered tools for the energy industry — geothermal, renewable, oil & gas, and nuclear. Our models are trained on real simulation data and validated against industry benchmarks.

• Platform: platform.forcex-ai.com
• Website: forcex-ai.com